Tagged "model-optimization"

279 articles tagged model-optimization, 11 February 2026 to 25 September 2026. Newest first.

• BottleCap AI Releases ThinkingCap-Qwen3.8-27B with 37% Fewer Thinking Tokens 25 September 2026

  A new specialized model variant optimizes Qwen3.8-27B by reducing inference thinking tokens by 37.2% with only marginal accuracy loss. This significantly reduces computational overhead for local deployments running reasoning workloads.

• Husky: Model-Specific Inference Engine Achieves 4.5x Speedup Over Apple MLX 24 September 2026

  A new inference engine optimised for Apple Silicon demonstrates dramatic performance improvements over existing solutions, achieving up to 4.5x faster inference than MLX for specific model architectures.

• Ollama 0.34.1 Stabilizes MLX Backend and GGUF Model Creation 16 September 2026

  Ollama v0.34.1 releases improved MLX memory handling for Apple Silicon, stabilizes GGUF creation workflows, and enhances repeat token detection for more reliable local inference.

• OpenBMB Releases MiniCPM5-2B as State-of-the-Art Open Model Under 4B Parameters 11 September 2026

  OpenBMB's MiniCPM5-2B achieves state-of-the-art performance for models under 4 billion parameters, making it ideal for on-device deployment scenarios with strict resource constraints. This release demonstrates significant progress in model efficiency without sacrificing capability.

• Cambricon Adapts DeepSeek-V4.1-Flash on vLLM Stack for Efficient Inference 11 September 2026

  Cambricon's Day-0 project successfully adapts DeepSeek-V4.1-Flash within the vLLM inference stack, demonstrating practical optimization of large open models for deployment. This work bridges advanced open models with production-grade serving infrastructure.

• Benchmarking Qwen 3.8 27B Quantizations: 4-Bit Holds Up, 1-Bit Collapses 9 September 2026

  Detailed quantization benchmarks for Qwen 3.8 27B revealing how 4-bit quantization maintains model quality while 1-bit approaches fail significantly.

• llama.cpp Optimizes DFlash Encoder with KV Cache Injection 31 August 2026

  Recent llama.cpp builds include performance improvements for DFlash models by fusing encoder operations into KV cache injection, reducing computational overhead for local inference.

• llama.cpp Build b10581 Adds DSpark Support for Faster Local Inference 22 August 2026

  The latest llama.cpp release includes native support for DSpark model optimization, enabling users to run DSpark-optimized models like LFM2.5 with maximum efficiency. This update extends llama.cpp's lead as the fastest local inference engine.

• Liquid AI Releases DSpark Version of Compact LFM2.5 Models with Up to 2.67x Speedup 22 August 2026

  Liquid AI releases optimized DSpark variants of their LFM2.5 models, achieving up to 2.67x inference speedup. These compact models are designed for on-device and edge deployment scenarios where latency and resource constraints are critical.

• Qwen3.8-27B: Running a Frontier-class Open Model on Your Local GPU 21 August 2026

  A comprehensive guide to deploying Qwen3.8-27B, a frontier-class open model, on consumer GPUs with practical optimization techniques for local inference.

• Gemma 4 Turns Ancient Laptops Into Dedicated Local LLM Inference Stations 20 August 2026

  How-To Geek reports on Gemma 4's efficiency improvements that enable capable local LLM inference even on older hardware. Gemma 4 represents a breakthrough in making modern language models viable for resource-constrained devices.

• Qwen3.8-27B Matches Claude Opus 4.6 on Coding, Runs on Consumer GPUs 19 August 2026

  Alibaba's Qwen3.8-27B model achieves performance parity with Claude Opus 4.6 on coding benchmarks while remaining deployable on consumer-grade GPUs, representing a major milestone for affordable local LLM inference.

• Hugging Face State of Open Models: Summer 2026 Observations 15 August 2026

  Hugging Face publishes comprehensive analysis of the open model landscape in Summer 2026, documenting trends in model optimization, deployment patterns, and ecosystem maturation for local LLM inference.

• DeepSeek V4 Flash Achieves 82.7% on Terminal-Bench 2.1 9 August 2026

  DeepSeek V4 Flash demonstrates strong benchmark performance with 82.7% accuracy on Terminal-Bench 2.1 using a public harness. This efficient model variant shows promise for local deployment scenarios requiring high capability with reasonable resource constraints.

• Optimizing Qwen 3.6 for Local Development: A Developer's Guide 7 August 2026

  A practical developer guide for optimizing the Qwen 3.6 model specifically for local development environments, covering configuration and performance tuning.

• Apple's Hardware Is Ready for On-Device AI and PrismML Just Delivered a Real Breakthrough 3 August 2026

  Apple's latest hardware capabilities combined with PrismML breakthroughs enable practical on-device AI inference, signaling mature support for local LLM deployment on iOS and macOS ecosystems.

• Tim Cook Called Apple's On-Device AI a 'Competitive Weapon' in Final Earnings Call as CEO 1 August 2026

  Apple's leadership emphasizes on-device AI as a strategic differentiator, signaling major investment in local inference capabilities. This reflects industry momentum toward edge deployment and privacy-first AI architectures.

• Samsung's Newest Foldable Phones Use Google's Gemini Nano 4 On-Device AI Model 31 July 2026

  Samsung has integrated Google's Gemini Nano 4 directly into its latest foldable phones for on-device AI processing. This mainstream adoption demonstrates the maturation of small, efficient models optimized for local inference on consumer hardware.

• OPPO Launches Xiaobu Next Beta, Debuts On-Device Multi-Agent System on Smartphones 27 July 2026

  OPPO has released a beta version of Xiaobu Next, an on-device multi-agent AI system that runs directly on smartphones without cloud connectivity. This represents a significant milestone in bringing advanced LLM capabilities to consumer mobile hardware.

• No Wi-Fi, No Data Transfer, Tablets Can Now Summarise Sensitive Documents Locally 26 July 2026

  Tablets can now process and summarize sensitive documents entirely on-device without requiring internet connectivity or data transfer. This advancement demonstrates practical deployment of LLMs on mobile hardware for enterprise document processing.

• From Foldables to Smart Glasses, Samsung's Galaxy AI Push Moves Beyond the Cloud 26 July 2026

  Samsung is shifting Galaxy AI capabilities from cloud-dependent processing to on-device edge inference across multiple device categories including foldables and smart glasses. This major OEM commitment signals mainstream adoption of local LLM deployment.

• AMD Advancing AI 2026: Enterprise AI Architecture Basics for Startup Founders 26 July 2026

  AMD is providing enterprise AI architecture guidance focused on practical deployment patterns. The content addresses foundational architecture decisions for startups building AI systems, including considerations for local and edge inference infrastructure.

• Multiverse Computing's CompactifAI Models Now Fully Compatible with Intel Xeon 6 Processors 24 July 2026

  All CompactifAI optimised models have achieved compatibility with Intel Xeon 6 processors, enabling efficient inference on enterprise server hardware and expanding deployment options for self-hosted local LLM infrastructure. This compatibility expands the practical deployment platforms for optimised models.

• Google's Gemma AI Runs Locally on a $300 Mini PC, and It Replaced ChatGPT 23 July 2026

  Google's Gemma model demonstrates practical feasibility of running capable local LLMs on ultra-budget hardware, showing that effective AI inference is now accessible to mainstream users without cloud dependency.

• AI Data Center Power Constraints Are the Real 2026 Bottleneck 20 July 2026

  Analysis of how power limitations in data centers are becoming the primary constraint for AI infrastructure, with implications for distributed and edge deployment strategies.

• Agentic Test Processes and LLM Benchmarks: Evaluating Local AI Agents 20 July 2026

  Dan Luu explores agentic test processes and LLM benchmarking methodologies, providing insights into how to properly evaluate language models in autonomous agent scenarios.

• Samsung Galaxy Watch 9 to Feature Snapdragon Wear Elite Chip: Report 19 July 2026

  Samsung's upcoming Galaxy Watch 9 is expected to include Qualcomm's new Snapdragon Wear Elite chip, enabling more sophisticated on-device AI capabilities on wearable devices.

• Qualcomm's Xu Hao: Agentic AI Phones Surge as On-Device AI Shifts from Passive Response to Proactive Service 19 July 2026

  Qualcomm executive highlights the shift toward agentic AI capabilities on mobile devices, moving beyond simple query-response patterns to proactive, autonomous service delivery on-device.

• Nvidia Showcases Nemotron Models for Japanese AI Development 17 July 2026

  Nvidia highlights its Nemotron model family's application in Japanese AI development, emphasizing locally-deployable language models optimized for specific regions and use cases.

• Google Demonstrates New On-Device AI Features for Pixel 10 17 July 2026

  Google has unveiled new on-device AI capabilities for the upcoming Pixel 10, showcasing advances in edge inference that run directly on mobile hardware without cloud connectivity. These features highlight the industry's momentum toward practical local LLM deployment on consumer devices.

• Google Gemma 4 Debuts for Pixel 10 With Powerful On-Device AI Features 16 July 2026

  Google has released Gemma 4, a new model family optimized for on-device inference on Pixel 10, demonstrating production-grade implementation of privacy-first AI. The model family represents important architectural improvements for resource-constrained edge deployment.

• Mira Murati's Thinking Machines Launches Open-Weight AI Model 16 July 2026

  Former OpenAI CTO Mira Murati's new venture, Thinking Machines, has released an open-weight AI model competing with NVIDIA's Nemotron. The model prioritizes efficiency and open deployment, expanding quality options for local LLM practitioners.

• Google expands on-device AI for Pixel phones with Gemma 4 15 July 2026

  Google brings its latest Gemma 4 model to Pixel devices with on-device optimization, expanding the availability of capable local LLMs on consumer hardware.

• Don't Sleep on BitNet (2025) 15 July 2026

  An exploration of BitNet technology and its implications for efficient local language model inference, highlighting how ultra-low-bit quantisation techniques can dramatically reduce model size and memory requirements.

• Google Pixel Implements Local AI for Screenshot Analysis With Privacy Controls 12 July 2026

  Google demonstrates on-device AI processing for Pixel screenshot features, keeping image analysis local while maintaining user privacy rather than routing data to cloud services.

• Edge AI Transformation Coming to Creative Production Workflows 6 July 2026

  Industry analysis shows edge AI is poised to reshape creative production, with on-device inference enabling real-time processing without cloud dependencies. Local LLMs will play a key role in this shift.

• Google Rolls Out Android 17 and Gemma 4 with Advanced On-Device AI 6 July 2026

  Google's latest Android 17 release integrates Gemma 4, bringing improved on-device AI capabilities optimized for local inference. The new features enable developers to deploy advanced language models directly on Android devices.

• Privatewhisper.ai: Private AI Voice Dictation Without Typing 29 June 2026

  Privatewhisper.ai enables on-device speech-to-text processing using local models, offering privacy-preserving voice dictation without sending audio to cloud servers.

• Giving AI Human-Like Memory Limits (3–7 Words) Could Improve Language Learning 24 June 2026

  Research from the Max Planck Institute reveals that constraining AI model memory to human-like limits may enhance language learning efficiency. This discovery has implications for optimizing local LLM training and inference under resource constraints.

• Boost Inference Performance up to 15x on NVIDIA Blackwell Using DFlash Speculative Decoding 24 June 2026

  NVIDIA introduces DFlash speculative decoding technique achieving up to 15x inference speedup on Blackwell GPUs, a major breakthrough for accelerating local LLM deployments on enterprise hardware.

• What else is included in the 'GGUF' file format used by llama.cpp for AI language models, besides weights? 21 June 2026

  An in-depth technical analysis of the GGUF format ecosystem, exploring the metadata, configuration, and structural components beyond model weights. Understanding GGUF is essential for practitioners working with llama.cpp and quantized model deployment.

• Qualcomm Launches Snapdragon START to Speed AI Smart Glasses to Market 20 June 2026

  Qualcomm's new Snapdragon START platform aims to accelerate edge AI deployment on smart glasses and mobile devices, providing optimized hardware for local LLM inference.

• Chrome Is Hiding a Free Local AI Chatbot on Your Computer 18 June 2026

  Google Chrome now includes a built-in local AI chatbot that runs directly on your machine without requiring cloud connectivity. This represents a significant shift toward edge inference in mainstream browsers.

• On-Device AI Market Projected to Reach $75.5 Billion by 2033 18 June 2026

  Market research predicts explosive growth in the on-device AI sector, driven by demand for real-time intelligence and privacy-first computing. The market is expected to expand significantly as edge inference becomes mainstream across consumer and enterprise applications.

• Apple Unveils AFM 3 Core Advanced with 20 Billion Parameters for On-Device AI 10 June 2026

  Apple introduced the AFM 3 Core Advanced architecture at WWDC26, featuring a 20 billion parameter model optimized for on-device inference. This represents a significant milestone in local LLM deployment on consumer hardware with architectural innovations to overcome memory constraints.

• A Cinematic Landing-Page Hero for 80 Cents (GPT Image 2 and Veo 3.1) 2 June 2026

  A cost-effective demonstration of generating cinematic video content for landing pages using recent image and video generation models, highlighting practical economics of modern generative AI.

• JetBrains Releases Mellum2: A 12B MoE Model for Fast, Specialized Tasks 2 June 2026

  JetBrains introduces Mellum2, a 12-billion parameter mixture-of-experts model designed for efficient local inference in multi-model AI pipelines. The model balances performance and resource consumption for on-device deployment scenarios.

• Why Chinese AI Labs Went Open and Will Remain Open 31 May 2026

  An examination of why leading Chinese AI laboratories have adopted open-source strategies and how this trend impacts the global LLM landscape and local deployment ecosystem.

• Liquid AI Launches Edge-Focused LFM2.5 Model to Power On-Device AI Agents 31 May 2026

  Liquid AI has released the LFM2.5 model specifically optimized for edge deployment and on-device AI agents. This new model represents a significant development for practitioners looking to run capable language models locally with reduced resource requirements.

• What Apple Knows About AI That Silicon Valley Won't Admit 31 May 2026

  An analysis of Apple's approach to on-device AI and the practical wisdom the company has gained from years of edge inference experience that challenges mainstream cloud-centric AI assumptions.

• Liquid AI Unveils Edge-Focused LFM2.5 Model for On-Device AI Agents 29 May 2026

  Liquid AI has introduced the LFM2.5 model specifically designed for edge deployment and local AI agents, offering optimized performance for resource-constrained environments.

• Google Launches Tiny Board for Running Gemma 3 Locally 29 May 2026

  Google has released a compact development board designed to run Gemma 3 models locally, making edge inference more accessible for developers and makers without requiring significant hardware investment.

• Alibaba Cloud Joins PyTorch Foundation as Platinum Member 28 May 2026

  Alibaba Cloud's elevation to PyTorch Foundation Platinum membership indicates major enterprise backing for the deep learning framework, with implications for distributed training and on-device optimization tooling.

• MediaTek Dimensity 8550 Shifts Focus to Gemini Nano V3 and On-Device AI on Phones 28 May 2026

  MediaTek's Dimensity 8550 processor emphasizes on-device AI capabilities optimized for Gemini Nano V3, advancing the smartphone landscape for local language model inference.

• The Anatomy of an LLM 28 May 2026

  A technical deep-dive into how large language models work internally, covering architecture, training, and inference fundamentals essential for understanding local deployment.

• OpenBMB Runs Local Agents with MiniCPM5-1B – Efficient LLM for Edge Deployment 27 May 2026

  OpenBMB demonstrates local agent execution using MiniCPM5-1B, an extremely efficient model optimized for on-device inference and agentic workflows.

• DeepSeek's Flagship V4 Pro Model Drops to 75% Lower Pricing, Increasing Competitive Pressure on Local Inference Economics 26 May 2026

  DeepSeek permanently reduced V4 Pro pricing by 75%, reshaping the cost-benefit analysis for developers deciding between cloud API usage and self-hosted local LLM deployment.

• Developer Switches from LM Studio to llama.cpp, Reports No Performance Downgrade 26 May 2026

  A developer shares their experience migrating from LM Studio to llama.cpp for local LLM inference, finding the lighter-weight tool delivers comparable performance with better resource efficiency.

• Show HN: An Open-Source Interactive AI Engineering Syllabus (1,100 Papers) 25 May 2026

  Community-driven curriculum curating 1,100 papers on AI engineering released as open-source resource. Valuable reference for understanding foundations of model optimization, deployment, and inference techniques.

• Maker Demonstrates Portable AI with Suitcase-Integrated Jetson Orin Setup 25 May 2026

  A maker successfully built a mobile AI assistant using NVIDIA's Jetson Orin, showcasing practical edge deployment potential for local models in portable form factors.

• Apple's 2026 AI Strategy Prioritizes On-Device Model Deployment 25 May 2026

  Apple is shifting its AI roadmap toward on-device model execution, signaling industry momentum toward privacy-preserving local inference.

• Gemma 4: A New Budget-Focused Model in Posit AI 25 May 2026

  Google releases Gemma 4, a new lightweight model optimized for budget-conscious local deployment scenarios. This addition to the Gemma family targets edge inference and resource-constrained environments.

• Redditor Successfully Runs 1 Trillion Parameter LLM Using Cheap Intel Optane DIMMs 24 May 2026

  A creative hardware hack demonstrates running a trillion-parameter LLM using affordable Intel Optane DIMM memory, achieving a breakthrough in cost-effective large model deployment. The approach opens new possibilities for running massive models on constrained budgets.

• Google Makes Gemini 3.5 Flash the Default AI Model for Billions of Users 22 May 2026

  Google's decision to make Gemini 3.5 Flash the default model for billions of users signals industry trends toward smaller, faster models optimized for on-device and edge inference. This shift has implications for local LLM development and deployment strategies.

• The Brain vs. Deep Learning Part I: Computational Complexity Analysis 22 May 2026

  A detailed analysis comparing computational complexity between biological brains and deep learning systems provides theoretical foundations for understanding efficiency trade-offs in model design and local deployment. This research is foundational for optimizing inference on resource-constrained devices.

• Google's Cormac Brick on Tiny LLMs for On-Device Agents 21 May 2026

  Google shares insights on deploying tiny language models optimized for on-device agents, offering practical perspectives on model size, latency, and autonomous decision-making at the edge.

• Meta Plans Agentic AI on Smartphones and Wearables by 2026 20 May 2026

  Meta Reality Labs outlines roadmap for deploying agentic AI systems directly on smartphones and wearables. The initiative aims to bring autonomous AI agents to consumer devices within the next two years.

• On-Device AI to Be in 80% of Wearables by 2032 19 May 2026

  Market research projects that on-device AI will become standard in 80% of wearables by 2032, driving demand for ultra-efficient models and hardware optimized for constrained environments. This trend indicates significant growth opportunities for local LLM deployment on edge devices.

• Ansede-static: Offline SAST Tool Demonstrates Value of Local AI Tools 18 May 2026

  New open-source static analysis tool achieving 98.8% CVE recall while running entirely offline. Exemplifies how local AI models can replace cloud-based security analysis with privacy-preserving alternatives.

• The Time Bomb Went Off: AI's All-You-Can-Eat Era Just Ended in Real Time 18 May 2026

  Cloud API pricing models are shifting away from subsidized unlimited access, making local LLM deployment increasingly economical. Market analysis of how API cost changes drive adoption of on-device inference.

• Local LLMs Enable Intelligent Smart Camera Control Without Cloud Dependency 18 May 2026

  A hands-on exploration demonstrates how local language models can power video doorbell intelligence and smart camera decision-making, eliminating latency and privacy concerns of cloud-based vision AI.

• A Cheap Fix That Saves the AI $400M Dollars a Year and Brings 4B People Online 17 May 2026

  An exploration of cost-effective infrastructure solutions with implications for understanding economic drivers behind local and edge LLM deployment at scale.

• Google Limits Gemini Intelligence to New Flagships—Hardware Requirements for Local Deployment 17 May 2026

  Google has unveiled Gemini Intelligence capabilities restricted to flagship devices, with extreme hardware requirements that limit deployment scope. This underscores the ongoing challenge of fitting capable AI models into accessible, consumer-level hardware.

• A Lo-Fi Rebellion Against A.I 17 May 2026

  An examination of a growing movement questioning uncritical AI adoption, with implications for understanding local LLM use cases and the demand for alternative, human-controlled approaches to AI systems.

• Local LLM Integration Enables Replacement of Paid Subscription Services 16 May 2026

  A practitioner demonstrates replacing three subscription-based applications by deploying a local language model with access to personal files, showcasing cost savings and privacy benefits.

• Orthrus Reshapes Economics of Local AI Inference with New Optimization Approach 16 May 2026

  Orthrus introduces breakthrough optimization techniques that make local AI inference economically viable for more use cases and deployment scenarios.

• AI/ML Benchmark Tool for Local LLM Inference and XGBoost Training 16 May 2026

  A new benchmarking tool has been released for measuring local LLM inference performance and XGBoost training across GPU and CPU hardware. This resource helps practitioners evaluate their on-device deployment setups and optimize inference performance.

• LLM temporal and causal reasoning research 15 May 2026

  New research repository exploring how local LLMs can improve temporal and causal reasoning capabilities, addressing a known limitation in current models. Understanding and improving these fundamental reasoning abilities is crucial for reliable local model deployment.

• Arm and Google Collaborate on On-Device AI Optimization Techniques 15 May 2026

  Arm and Google have published guidance on accelerating on-device AI inference, focusing on optimization strategies for edge devices and resource-constrained environments. The collaboration provides practical approaches for deploying LLMs efficiently on mobile and embedded systems.

• Open-Source Local LLM Emerges as Viable Cloud AI Competitor 15 May 2026

  A recent analysis demonstrates that open-source local LLMs now offer competitive performance with cloud-based AI services in many use cases. The findings highlight the maturing landscape of on-device inference and cost advantages of self-hosted solutions.

• Chrome Automatically Downloads 4GB AI Model for Local Processing 14 May 2026

  Google Chrome now automatically downloads a 4GB on-device AI model to support native AI features, with implications for local inference standards and user privacy. Users can disable the automatic download if preferred.

• Local LLM Persistent Context Prevents Repetitive Mistakes 14 May 2026

  A practitioner shares how implementing persistent context in their local LLM deployment significantly improved response consistency and reduced recurring errors. This technique enhances model performance without requiring model retraining or hardware upgrades.

• Running a Local LLM on a 12-Year-Old Raspberry Pi 13 May 2026

  A practical guide demonstrating how to successfully run local LLMs on legacy hardware, proving that edge inference is achievable even on severely resource-constrained devices like the original Raspberry Pi.

• MDL: Endless Visual Novel Engine Powered by AI 11 May 2026

  MDL showcases an AI-powered visual novel engine that leverages local inference for game content generation. This demonstrates creative applications of on-device LLMs in interactive entertainment.

• One LM Studio Setting Change Makes Local LLMs Competitive With Cloud Models 11 May 2026

  A simple configuration adjustment in LM Studio dramatically improves local LLM performance, making self-hosted inference viable for production workloads previously requiring cloud APIs. This discovery highlights how software optimization can rival hardware improvements.

• LibreOffice 26.4 Beta Integrates Local AI Writing Features 10 May 2026

  LibreOffice's latest beta introduces integrated AI writing capabilities, with potential for local model support in office productivity workflows.

• Qwen3-Coder-Next Local Deployment: Complete Developer Guide for 2026 10 May 2026

  A comprehensive guide for deploying Qwen3-Coder-Next, a state-of-the-art coding model optimized for local environments. The guide covers setup, configuration, and practical deployment strategies for developers.

• Small On-Device AI Model Beats Claude Sonnet 4.5 and GPT-5 10 May 2026

  A newly optimized on-device AI model demonstrates performance that exceeds leading cloud-based models on specific benchmarks. This breakthrough challenges assumptions about model size and cloud superiority for local deployment.

• DistillFast: AI Cost Optimization Tool for Model Efficiency 10 May 2026

  A new cost optimization tool focused on reducing computational overhead for AI inference, relevant for practitioners looking to maximize efficiency in local deployments.

• How I Used a Local LLM to Organize the Store on My NAS 9 May 2026

  A practical case study demonstrating how local LLMs can be deployed on Network Attached Storage systems for practical applications like file organization and metadata management without cloud connectivity.

• How to Run LLMs Locally on Your Laptop for Free: A Beginner's Guide 9 May 2026

  A comprehensive beginner's guide covering the fundamentals of running language models locally without cloud dependencies, including tools, hardware requirements, and practical setup instructions.

• Perplexity Brings On-Device AI Workflow to Macs with 'Personal Computer' Feature 8 May 2026

  Perplexity has launched an on-device AI workflow for macOS that brings privacy-preserving inference capabilities directly to users' machines. This represents a significant shift toward practical, privacy-first local LLM deployment on consumer hardware.

• Nota AI Partners with Mobilint to Accelerate On-Device AI on Domestic NPU Infrastructure 7 May 2026

  Nota AI has announced a strategic partnership with Mobilint focused on optimizing on-device AI deployment using Neural Processing Units (NPUs). This collaboration aims to commercialize AI optimization technology for domestic NPU infrastructure.

• Sarvam Edge: Indian-Built AI Models Run Offline on Phones and Laptops Without Internet 6 May 2026

  Sarvam AI released Sarvam Edge, a suite of models specifically designed for on-device deployment on smartphones and laptops without internet connectivity. This represents a significant step forward in making practical, localized AI accessible across diverse hardware.

• Google Accelerates Gemma 4 Inference Speed 3x With Multi-Token Prediction Drafters 6 May 2026

  Google announced significant performance improvements for Gemma 4 through multi-token prediction drafters, achieving 3x faster inference. This optimization technique is directly applicable to local LLM deployments and represents a major breakthrough in edge inference efficiency.

• A 49-Line Physics Classifier That Beats kNN on 76% of Benchmarks 5 May 2026

  A minimal, efficient physics classifier demonstrates that simple, optimized algorithms can outperform traditional machine learning approaches on standard benchmarks with dramatically reduced code complexity.

• Supercharging LLM Inference on Google TPUs: Achieving 3X Speedups With Diffusion-Style Speculative Decoding 5 May 2026

  Google researchers have demonstrated 3x inference speedups on TPUs using diffusion-style speculative decoding, a novel optimization technique that could influence local inference strategies. The breakthrough shows how advanced decoding methods can dramatically reduce latency on specialized hardware.

• Google's Gemma 4 Could Put Powerful AI on Your Phone and Laptop 5 May 2026

  Google is advancing on-device AI capabilities with Gemma 4, a model family optimized for edge deployment on consumer devices. This release signals a major push toward bringing sophisticated language models to phones and laptops without cloud dependencies.

• Google Explains Why AICore Storage Requirements Are Increasing on Android 4 May 2026

  Google provides transparency about the expanding storage footprint of AICore, its on-device AI runtime for Android, explaining the tradeoffs between capability and storage size.

• Major Smartphone Brands Introduce Advanced On-Device AI Features 4 May 2026

  Leading smartphone manufacturers are rolling out sophisticated on-device AI capabilities, signaling broad industry momentum toward local model inference on mobile hardware.

• Anker's Thus Chip Puts AI On-Device, Promising Faster Responses And Better Privacy 4 May 2026

  Anker introduces the Thus chip, a dedicated hardware accelerator designed to run AI models entirely on-device with improvements in response latency and privacy preservation.

• I Put a Local LLM on My Phone and Stopped Needing Cloud AI for Most Tasks 3 May 2026

  Practical demonstrations show that modern optimized language models can run efficiently on smartphones, eliminating cloud API dependency for many everyday AI tasks. Mobile local inference offers privacy, offline availability, and reduced latency for real-world applications.

• Running a Serious AI Model on a Consumer GPU Just Got Easier and That Matters More Than the Benchmark 3 May 2026

  Recent advances in optimization techniques and frameworks have made it significantly easier to run production-quality large language models on consumer-grade GPUs, democratizing access to capable local AI inference. Performance improvements go beyond raw speed gains to include better memory efficiency and developer experience.

• Study: AI Models That Consider User Feelings Are More Likely to Make Errors 2 May 2026

  Research reveals that adding empathy or emotional responsiveness to AI models reduces factual accuracy, with important implications for deploying local LLMs in critical applications. The findings suggest developers should optimize for task-specific accuracy rather than alignment for all use cases.

• Anker's New 'Thus' Chip Brings 150x AI Power to Earbuds 2 May 2026

  Anker has announced a specialized AI chip for earbuds that dramatically increases on-device processing capability, enabling local inference on ultra-constrained hardware.

• Home Assistant's Local LLM Support Outperforms Gemini for Home Automation 1 May 2026

  Home Assistant's integrated local LLM capabilities now outperform Google's Gemini for smart home tasks, demonstrating the practical advantages of on-device inference for privacy-critical applications.

• Estimating Black-Box LLM Parameter Counts via Factual Capacity 30 April 2026

  New methodology for determining LLM model size without access to weights, enabling better deployment decisions and benchmarking for local inference scenarios.

• How Much "Brain Damage" Can an LLM Tolerate? 30 April 2026

  Research explores LLM resilience to model degradation, weight pruning, and parameter corruption—critical insights for optimizing models for edge and resource-constrained deployments.

• Stop Guessing: Open-Source Tool Predicts Which Local LLMs Run on Your PC 28 April 2026

  A new open-source diagnostic tool helps practitioners quickly determine which language models will run efficiently on their specific hardware without trial and error. This addresses a major pain point in local LLM adoption.

• Google's Gemma 4: Powerful AI Models Optimized for Your Phone and Laptop 28 April 2026

  Google introduces Gemma 4, a new generation of AI models specifically engineered for efficient on-device inference on phones and laptops. These models represent a major step forward in bringing capable language models to edge devices without cloud dependencies.

• Economic Implications of AI Adoption: Why Local Deployment Matters for Cost Control 28 April 2026

  An examination of the economic disparities in AI access and adoption, with implications for cost-conscious organizations considering local LLM deployment.

• Google's Gemma 4 Could Put Powerful AI on Your Phone and Laptop 26 April 2026

  Google's new Gemma 4 model is designed for efficient on-device deployment across phones and laptops, bringing capable inference to edge devices without cloud dependency.

• Run a Local LLM Server on Raspberry Pi with Remote Access Capabilities 25 April 2026

  A practical demonstration of deploying inference-optimized LLMs on Raspberry Pi hardware with remote accessibility, proving that edge AI inference doesn't require expensive equipment. This enables truly distributed, cost-effective local AI deployments.

• Google's Gemma 4 Brings Powerful On-Device AI to Phones and Laptops 25 April 2026

  Google announces Gemma 4, an optimized model family designed specifically for efficient on-device inference on consumer hardware. This release demonstrates the industry-wide shift toward practical edge AI deployment.

• How to Make Sense of AI 24 April 2026

  CommonCog publishes a comprehensive guide to understanding AI systems, providing essential context for practitioners evaluating and deploying local LLMs effectively.

• I Replaced My Local LLM With a Model Half Its Size and Got Better Results 24 April 2026

  Case study demonstrating that model size isn't the only factor determining performance—proper quantization, fine-tuning, and hardware matching can yield superior results with significantly smaller models.

• Show HN: We built an OCR server that can process 270 dense images/s on a 5090 23 April 2026

  A high-performance OCR inference server achieving 270 dense images per second on a single GPU, demonstrating practical edge inference optimization techniques.

• Llama 4 Scout on MLX: The Complete Apple Silicon Guide (2026) 23 April 2026

  An updated guide for running Llama 4 Scout models on Apple Silicon using MLX, covering optimization techniques and practical deployment patterns for macOS-based local LLM inference.

• Sarvam Edge: India's Offline AI Model Runs on Phones and Laptops Without Internet 22 April 2026

  Sarvam AI has released Edge, an AI model specifically designed for on-device inference on mobile phones and laptops that operates entirely offline. The model represents a regional approach to practical edge deployment optimized for Indian languages and use cases.

• Gemma 4 Just Replaced My Whole Local LLM Stack 19 April 2026

  Google's Gemma 4 model is making waves in the local LLM community as users report it outperforming their entire previous inference stacks. The model appears to deliver significant improvements in performance and efficiency for on-device deployment.

• Bonsai 1.7B in the Browser: A 290MB 1-bit LLM on WebGPU 16 April 2026

  Bonsai, a 1.7B parameter model quantized to 1-bit, now runs directly in web browsers via WebGPU at just 290MB. This breakthrough demonstrates extreme quantization techniques making capable language models viable for edge inference without server infrastructure.

• Building a Voice AI Wearable in a Casio F91W with Whisper and BLE 16 April 2026

  A developer successfully embedded voice AI capabilities into a classic Casio F91W watch using an nRF52840 microcontroller, Whisper speech-to-text, and Bluetooth Low Energy. This demonstrates practical on-device speech processing on severely constrained hardware.

• DotLLM – Building an LLM Inference Engine in C# 15 April 2026

  A new LLM inference engine implementation in C# provides .NET developers with native capabilities for running language models locally. This expands the ecosystem of local inference frameworks beyond Python-dominant tooling.

• Running Gemma 4 on an iPhone 13 Pro 15 April 2026

  A developer successfully demonstrates running Google's Gemma 4 model directly on iPhone 13 Pro hardware using LiteRTLM-Swift. This showcases practical on-device inference capabilities for modern mobile devices without cloud dependencies.

• SigMap – Shrink AI Coding Context 97% with Auto-Scaling Token Budget 15 April 2026

  SigMap introduces an auto-scaling token budget system that reduces AI coding context by 97%, enabling more efficient local model inference for code generation and analysis tasks. This performance optimization is critical for running models on memory-constrained devices.

• Google's Gemma 4 Brings Game-Changing Performance to Local Laptop Inference 15 April 2026

  Google and NVIDIA collaborate to optimize Gemma 4 for on-device laptop deployment, enabling efficient local inference without cloud dependencies. This advancement demonstrates significant progress in making capable language models accessible for personal computing.

• Qwen 3.5 Small – On-Device Multimodal Models Released 14 April 2026

  Alibaba's Qwen team has released Qwen 3.5 Small, a new multimodal model optimized for on-device inference. This lightweight model enables local deployment of vision and language capabilities without cloud dependencies.

• MiniMax M2.7 Open-Sources Globally as Industry's First Self-Improving Model 13 April 2026

  MiniMax has open-sourced its M2.7 model globally, introducing a self-improving capability that allows the model to optimize its own performance. This release significantly expands options for local deployment of sophisticated, autonomously-improving language models.

• Learn LLM Internals 13 April 2026

  A comprehensive GitHub repository documenting the internal mechanics of large language models, providing developers with deep knowledge necessary for optimizing local deployments. Essential reference material for understanding how to tune and optimize models running on limited hardware.

• MiniMax-M2.7 Delivers Exceptional Performance on Consumer Hardware 13 April 2026

  MiniMax-M2.7 benchmarks show strong throughput (127.7 tok/s on dual RTX PRO 6000 Blackwell) and efficient VRAM utilization, positioning it as a practical alternative to larger models for resource-constrained deployments.

• The Best Local AI Model for Home Assistant Isn't Always the Biggest One 12 April 2026

  A practical guide examining model selection for Home Assistant, revealing how optimal performance requires balancing model capability with hardware constraints rather than simply choosing the largest available model.

• MiniMax M2.7 Is Now Open Source 12 April 2026

  MiniMax releases M2.7, an agentic model now available as open source, expanding options for local deployment of capable reasoning models without cloud dependencies.

• Unsloth Completes Comprehensive MiniMax M2.7 GGUF Quantization Suite 12 April 2026

  Unsloth has finished quantizing MiniMax M2.7 across the full range of GGUF quantization levels from 1-bit to BF16, providing practitioners with optimized variants for every hardware configuration from edge devices to high-end systems.

• On-Device AI: Achieving Powerful AI Capabilities Without Internet Connectivity 12 April 2026

  An analysis of how modern on-device AI systems enable sophisticated AI capabilities entirely locally, examining the technical approaches and practical implications for truly disconnected deployment scenarios.

• Google Gemma 4 Delivers Exceptional Speed and Accuracy for Local Inference 12 April 2026

  Early adopters report that Google's Gemma 4 model runs with remarkable speed comparable to 4-9B parameter models while maintaining accuracy levels reminiscent of early Gemini releases, making it a compelling option for resource-constrained local deployments.

• Google's Gemini Nano 4 Offers Faster, Smarter Local Inference Capabilities 11 April 2026

  Google's latest Gemini Nano 4 model brings improved performance and speed for on-device AI inference. The model represents a significant step forward for local LLM deployment on edge devices and mobile platforms.

• Qualcomm Snapdragon XR Powers Next-Generation AI Glasses with Local Inference 11 April 2026

  Qualcomm's expansion of its XR collaboration with Snap demonstrates commitment to embedding powerful on-device AI in wearable hardware. The Snapdragon XR chip will enable local processing of AI workloads on upcoming AR glasses.

• AI PC Market Projected to Reach $235B by 2032, Driven by On-Device Computing Adoption 11 April 2026

  Market analysis predicts explosive growth in AI-enabled PCs powered by on-device inference capabilities. The trend reflects growing enterprise and consumer demand for local AI computing without cloud dependencies.

• ASUS ExpertBook P1 Integrates On-Device AI for Enterprise Collaboration 11 April 2026

  ASUS launches the ExpertBook P1 with integrated on-device AI collaboration tools, bringing local inference to enterprise computing. The laptop demonstrates practical implementation of privacy-preserving AI features for professional workflows.

• Community Reverse Engineers Gemma 4 Multi-Token Prediction Capability 10 April 2026

  Researchers have extracted Gemma 4 model weights and discovered multi-token prediction (MTP) functionality, launching a collaborative effort to understand and implement this capability for local models.

• Building Offline AI Companions on Severely Constrained Hardware (8GB RAM) 10 April 2026

  A practical case study demonstrates deploying local LLMs for accessibility applications with extreme hardware constraints, addressing real-world use cases where cloud deployment is infeasible.

• Run Qwen3.5 on an Old Laptop: A Lightweight Local Agentic AI Setup Guide 9 April 2026

  KDnuggets publishes a practical guide demonstrating how to run Qwen3.5 with agentic AI capabilities on resource-constrained hardware, making advanced local inference accessible to resource-limited environments.

• Running a 1.7B Parameters LLM on an Apple Watch 9 April 2026

  A developer successfully deployed a 1.7 billion parameter language model on an Apple Watch, demonstrating extreme edge inference capabilities on ultra-constrained wearable hardware.

• I Replaced My Local LLM With a Model Half Its Size and Got Better Results — and It Wasn't About the Parameters 9 April 2026

  A detailed account of how switching to a smaller, better-optimized model outperformed a larger predecessor on local hardware, challenging assumptions about model scaling and practical performance.

• Google AI Edge Gallery Showcases Offline Inference with Gemma 4 8 April 2026

  Google has launched the AI Edge Gallery application demonstrating practical use cases for offline inference with Gemma 4 on iOS and Android, including offline dictation and on-device AI features without internet connectivity.

• Lenovo Korea Launches AI-Powered Industrial Edge Solutions 6 April 2026

  Lenovo Korea has introduced artificial intelligence-based industrial edge solutions targeting manufacturing and enterprise environments. The products enable real-time AI inference at the edge without cloud connectivity dependencies.

• Show HN: Lightweight LLM Tracing Tool with CLI 6 April 2026

  A new open-source LLM tracing tool providing command-line observability for local language model deployments, helping developers debug and monitor inference pipelines.

• Google AI Edge Gallery Tops App Store Charts with On-Device Gemma 4 6 April 2026

  Google's AI Edge Gallery app has entered the App Store top 10, demonstrating mainstream adoption of on-device Gemma 4 models. The app enables users to run Google's latest locally-optimized LLM directly on their devices.

• NVIDIA and Google Optimize Gemma 4 AI Models for Local RTX Deployment 4 April 2026

  NVIDIA and Google have collaborated to optimize Gemma 4 models specifically for NVIDIA RTX GPUs, enabling high-performance local inference. The optimization work ensures efficient utilization of consumer and professional GPUs for on-device AI workloads.

• Google Launches Gemma 4 For Advanced On-Device AI 4 April 2026

  Google has released Gemma 4, an open model family designed for on-device AI inference across phones, tablets, and GPUs. The new models target efficient local deployment with improved capabilities for edge computing scenarios.

• AMD Rolls Out Gemma 4 Model Support Across Full Range of GPUs & CPUs 4 April 2026

  AMD has announced comprehensive support for Gemma 4 across its entire lineup of GPUs and CPUs, enabling local inference on AMD-based systems. The support extends from consumer Ryzen processors to professional EPYC servers and RDNA GPUs.

• VRAM Optimization Technique Cuts Gemma 4 Memory Usage by 3x 3 April 2026

  A simple llama.cpp parameter adjustment (-np 1) significantly reduces Sliding Window Attention cache VRAM requirements for Gemma 4, enabling deployment on systems with limited GPU memory.

• Qwen 3.6-Plus Released 2 April 2026

  Alibaba releases Qwen 3.6-Plus, a new model optimized for local deployment with improved performance characteristics for on-device inference.

• SmolLM2-360M Running on Samsung Galaxy Watch 4 with 74% Memory Reduction 2 April 2026

  Developer optimizes llama.cpp to run language models on smartwatches, achieving 74% RAM reduction through memory model improvements and reducing peak usage from 524MB to practical levels.

• TurboQuant Enables Qwen 3.5-27B on 16GB Consumer GPUs 2 April 2026

  Advanced quantization technique TurboQuant achieves near-Q4_0 quality at 10% smaller size, allowing high-performance models to fit on consumer-grade graphics cards.

• ByteShape Releases Qwen 3.5 9B Quantisations with Hardware-Matched Tuning Guide 1 April 2026

  ByteShape has released optimised GGUF quantisations of Qwen 3.5 9B with a comprehensive guide for selecting the best quantisation level for specific hardware. The resource includes comparative benchmarks against other popular quantisation approaches, enabling practitioners to make informed deployment decisions.

• Does RAG Help AI Coding Tools? 31 March 2026

  Analysis examining whether Retrieval-Augmented Generation actually improves code generation quality in AI coding assistants and local deployment scenarios.

• Local AI didn't replace my subscriptions, but it did take over these 6 tasks 31 March 2026

  A practical analysis of which specific workflows and tasks are most effective for local AI tools, helping practitioners identify high-impact use cases for self-hosted deployment.

• Running AI on a Raspberry Pi, Part 2: Running AI on a Pi in Under 5 minutes 31 March 2026

  A practical guide demonstrating how to deploy and run AI models on Raspberry Pi hardware in minimal time, making edge inference accessible to developers and hobbyists.

• DeepSeek V3 Complete Guide: Deploy and Optimize Local AI in 2026 30 March 2026

  A comprehensive guide for deploying and optimizing DeepSeek V3 for local inference, covering deployment strategies and optimization techniques for on-device AI applications.

• OLED Emerges as the Display Standard for Energy-Efficient AI Systems 29 March 2026

  As on-device AI inference becomes power-critical, OLED display technology is positioning itself as a key efficiency component in integrated AI systems, particularly for battery-constrained devices.

• DaVinci-MagiHuman: Open-Source AI Model for Realistic Video Generation 29 March 2026

  An open-source video generation model optimized for local inference, enabling developers to generate realistic videos on consumer hardware without cloud dependencies.

• TurboQuant KV Cache Compression Achieves 22.8% Faster Decoding at 32K Context 28 March 2026

  Google's TurboQuant compression method has been successfully integrated into llama.cpp, enabling 4.6x KV cache compression and 22.8% decode speedup at 32K context length by skipping 90% of dequantization work. This breakthrough makes long-context inference practical on consumer hardware like MacBook Air M4.

• Samsung Galaxy Book6 Series Brings Intel Core Ultra Chips for On-Device LLM Inference 28 March 2026

  Samsung's new Galaxy Book6 laptop series launched in India with Intel Core Ultra processors, targeting on-device AI capabilities and local LLM deployment on consumer hardware with improved neural processing performance.

• Quantization Reveals Outliers Impacting LLM Accuracy 27 March 2026

  Research reveals how outlier values in model weights and activations significantly impact accuracy when applying quantization to large language models. Understanding outlier handling is critical for effective model compression.

• This Wearable Runs an On-Device AI With 2-Week Battery Life 27 March 2026

  A new wearable device demonstrates practical on-device AI inference with exceptional battery efficiency, running for two weeks on a single charge. This showcases the feasibility of edge AI on severely resource-constrained devices.

• TurboQuant Benchmarked in Llama.cpp: Google's Extreme Compression Research Tested in Practice 27 March 2026

  Community members benchmarked Google's TurboQuant extreme compression technique within llama.cpp, providing practical performance data on the quantisation method. Results show how the research translates to real-world inference speed and memory usage improvements.

• Coding Implementation to Run Qwen3.5 Reasoning Models Distilled With Claude-Style Thinking Using GGUF and 4-Bit Quantization 27 March 2026

  A new implementation enables running distilled Qwen3.5 reasoning models with 4-bit quantization and GGUF format, making advanced reasoning capabilities accessible on consumer hardware. This combines distillation, quantization, and standardized formats for practical local deployment.

• Hold on to Your Hardware: Implications for Local LLM Deployment 27 March 2026

  An article examining hardware longevity and sustainability raises important considerations for practitioners investing in local inference infrastructure.

• Qwen 3.5 27B Achieves 1.1M Tokens/Second on B200 GPUs with Optimized vLLM Config 27 March 2026

  A developer optimized Qwen 3.5 27B to reach 1.1 million tokens per second on 96 B200 GPUs using vLLM, with detailed configurations and all settings published on GitHub. Key optimizations included distributed parallelism, reduced context windows, FP8 KV cache, and speculative decoding.

• RF-DETR Nano and YOLO26 Enable On-Device Object Detection on Smartphones 26 March 2026

  Researchers have demonstrated RF-DETR Nano and YOLO26 running object detection and instance segmentation on mobile phones entirely on-device, with no cloud API calls or external dependencies.

• Google's TurboQuant: The Unsexy AI Breakthrough Worth Watching 26 March 2026

  Google introduces TurboQuant, a quantization technique that enables efficient local LLM deployment by reducing model size and computational requirements without significant accuracy loss.

• Ultra-Large 400B-Class LLM Runs on iPhone in Test 25 March 2026

  A 400B-parameter language model has been successfully demonstrated running on an iPhone, marking a significant breakthrough in on-device inference capabilities. This achievement suggests that ultra-large models can now fit and execute on consumer mobile devices through advanced optimization techniques.

• Running an Open-Weight LLM Locally on an Apple Watch 25 March 2026

  A developer demonstrates successfully running an open-weight LLM directly on Apple Watch hardware, pushing the boundaries of edge inference on ultra-constrained devices.

• .APKs Are Just .ZIPs: Semi-Legally Hacking Software for Orphaned Hardware 25 March 2026

  A video explores reverse-engineering and modifying Android APKs to run on legacy devices, with techniques applicable to deploying inference engines on older hardware.

• LLM Neuroanatomy II: Modern LLM Hacking and Hints of a Universal Language 24 March 2026

  A deep technical exploration of LLM internals, examining how modern language models work at a fundamental level and uncovering potential universal patterns in their representations.

• Alibaba Commits to Continuous Open-Sourcing of Qwen and Wan Models 23 March 2026

  Alibaba has publicly committed to ongoing open-source releases of new Qwen and Wan models, reinforcing their position as a major contributor to the local LLM ecosystem. This commitment ensures continued availability of high-quality open-weight models for on-device deployment.

• Qwen 3.5 Models: Optimal Settings and Reduced Overthinking Configuration 23 March 2026

  Community exploration of Qwen 3.5 (35B and 27B) model settings and prompts reveals configurations that minimize overthinking behavior and excessive reasoning token usage. These practical optimizations help practitioners maximize output quality and inference speed.

• How to Build a Self-Hosted AI Server with LM Studio: Step-by-Step Guide 23 March 2026

  A comprehensive tutorial walks through deploying a self-hosted AI inference server using LM Studio, providing practical guidance for local LLM deployment.

• Powerful AI Search Engine Built on Single GeForce RTX 5090 23 March 2026

  An enthusiast successfully deployed a fully-featured AI search engine on a single GeForce RTX 5090 GPU, demonstrating the viability of complex local inference workloads on consumer hardware.

• Llama 8B Matches 70B Performance on Multi-Hop QA Using Structured Prompting 22 March 2026

  Structured prompting techniques with Graph RAG enable smaller Llama 8B models to match 70B model performance on complex multi-hop question answering without fine-tuning. Research reveals reasoning, not retrieval, is the actual bottleneck.

• Cursor's Composer 2 Model Analysis – Fine-Tuned Variant of Kimi K2.5 20 March 2026

  Community investigation reveals that Cursor's Composer 2 model appears to be based on Kimi K2.5 with reinforcement learning fine-tuning. This insight provides valuable intelligence about model adaptation techniques for local development environments.

• AI's Impact on Mathematics Analogous to Car's Impact on Cities 20 March 2026

  Mathematician Terence Tao shares perspective on how AI fundamentally reshapes mathematical practice and discovery, comparable to urban transformation. This philosophical analysis has implications for how local LLMs should be optimized for knowledge work.

• Ultra-Compact 28M Parameter Models Show Promise for Specialized Domain Tasks 20 March 2026

  Experimental work with tiny 28M parameter models fine-tuned on specific domains (like business email) reveals viable pathways for training task-specific models that run on extremely resource-constrained devices.

• Qwen 3.5 Emerges as Top Performer for Local Deployment with Extensive Quantization Options 20 March 2026

  Qwen 3.5 is establishing itself as a highly versatile model for local inference, with community members successfully creating dozens of custom quantizations and sharing best practices across different inference engines and hardware configurations.

• Community Converges on Optimal KV Cache Quantization Strategies for Qwen 3.5 Models 20 March 2026

  The local LLM community is establishing practical guidelines for KV cache quantization with Qwen 3.5, balancing memory savings against accuracy loss to optimize inference on consumer hardware.

• NVIDIA Nemotron Cascade 2 30B Delivers 120B-Class Performance in Compact Form Factor 20 March 2026

  NVIDIA's new Nemotron Cascade 2 30B achieves competitive performance with models 4x larger on math and code benchmarks, offering excellent efficiency for local deployment on resource-constrained hardware.

• MiniMax-M2.7: New Compact Model Announced for Local Deployment 18 March 2026

  MiniMax has announced the M2.7 model, generating interest in the community regarding its potential multimodal capabilities and suitability for local inference workloads.

• Local Qwen Models Master Browser Automation Through Iterative Replanning 17 March 2026

  Demonstration shows small local Qwen models (8B + 4B) dramatically improve browser automation accuracy by adopting a step-by-step replanning approach rather than generating full multi-step plans upfront.

• Researcher Discovers Universal "Danger Zone" in Transformer Model Architecture at 50% Depth 17 March 2026

  Experimental layer surgery across six different model architectures reveals a critical vulnerability at approximately 50-56% model depth where layer duplication consistently degrades performance, offering new insights into transformer architecture optimisation.

• Run LLMs Locally with Llama.cpp 17 March 2026

  A practical guide on leveraging llama.cpp for efficient local LLM inference, demonstrating how to optimize model performance on consumer hardware without cloud dependencies.

• Qwen 3.5 4B Outperforms Nvidia Nemotron 3 4B in Local Benchmarks 17 March 2026

  Community benchmarking reveals that Qwen 3.5 4B consistently outperforms Nvidia's newly released Nemotron 3 4B across demanding custom tests, challenging expectations for the Nemotron family.

• Practical Fix for Qwen 3.5 Overthinking in llama.cpp 16 March 2026

  Community members share techniques to mitigate Qwen 3.5's verbose internal reasoning loops, offering practical optimization strategies for controlling model behavior in local inference environments.

• Nota Added to Three Technology and Growth ETFs in a Row – Market Recognition for AI Efficiency 16 March 2026

  Nota's inclusion in multiple ETFs reflects investor confidence in neural network optimization technology. This signals market validation for quantization and efficiency innovations critical to local LLM deployment.

• OpenClaw Isn't the Only Raspberry Pi AI Tool—Here Are 4 Others You Can Try This Week 16 March 2026

  A survey of practical AI tools optimized for Raspberry Pi and other edge devices demonstrates the growing ecosystem of lightweight models and frameworks for constraint-based inference.

• StepFun Releases SFT Dataset Used to Train Step 3.5 Flash for Community Fine-Tuning 15 March 2026

  StepFun has open-sourced the supervised fine-tuning dataset behind Step 3.5 Flash, enabling local practitioners to understand, reproduce, and fine-tune efficient LLMs. This transparency advance the state of reproducible local LLM development.

• India's Mobile-First AI Strategy Could Accelerate Local Inference Adoption in Emerging Markets 15 March 2026

  India's playbook for mobile-first technology adoption offers lessons for democratizing AI inference in resource-constrained environments through local deployment.

• Local LLMs on Apple Silicon Mac 2026: M1 M2 M3 Guide 14 March 2026

  A comprehensive guide from SitePoint covering the latest techniques and models optimized for running local LLMs on Apple Silicon Macs in 2026. Essential reading for macOS users seeking practical deployment strategies.

• Show HN: Detect When an LLM Silently Changes Behavior for the Same Prompt 12 March 2026

  A new tool enables monitoring and detecting when LLMs silently alter their responses for identical prompts, addressing a critical reliability concern for production deployments.

• Qwodel – An Open-Source Unified Pipeline for LLM Quantization 12 March 2026

  Qwodel is a new open-source tool that provides a unified pipeline for LLM quantization, simplifying the process of reducing model size and improving inference speed for local deployment.

• Quantization Explained: Q4_K_M vs AWQ vs FP16 for Local LLMs 12 March 2026

  An in-depth technical guide comparing major quantization formats used in local LLM deployment, covering trade-offs between model size, inference speed, and quality.

• The $1,500 Local AI Setup: DeepSeek-R1 on Consumer Hardware 12 March 2026

  A comprehensive guide demonstrating how to deploy DeepSeek-R1 reasoning models on consumer-grade hardware for under $1,500, making advanced local inference accessible to individual developers.

• Nvidia Releases Nemotron 3 Super: 120B MoE Model for Local Deployment 12 March 2026

  Nvidia has released Nemotron 3 Super, a 120B mixture-of-experts model with only 12B active parameters, designed as an open-source alternative for agentic reasoning tasks. The hybrid Mamba-Transformer architecture offers competitive performance with reduced computational requirements.

• Simple Layer Duplication Technique Achieves Top Open LLM Leaderboard Performance 11 March 2026

  Researchers demonstrate that duplicating middle layers in Qwen2-72B without modifying weights produces state-of-the-art benchmark results, challenging conventional understanding of model optimization.

• Sarvam Open-Sources 30B and 105B Reasoning Models 11 March 2026

  Indian AI startup Sarvam has released open-source reasoning models in 30B and 105B parameter sizes, providing locally-deployable alternatives for reasoning tasks without reliance on proprietary APIs.

• Fine-Tuned Qwen SLMs (0.6–8B) Demonstrate Competitive Performance Against Frontier LLMs on Specialized Tasks 10 March 2026

  A systematic benchmarking study shows that properly fine-tuned Qwen3 small language models can match or exceed the performance of frontier LLMs like GPT-5 and Claude on narrowly-scoped tasks, validating the viability of local model specialization strategies.

• Qwen 3.5 Small Expands On-Device AI to Phones and IoT with Offline Support 9 March 2026

  Alibaba's Qwen 3.5 Small model brings efficient LLM inference to mobile devices and IoT hardware with full offline capabilities. This lightweight model expansion enables practical on-device deployment where connectivity and compute resources are severely constrained.

• How to Run Your Own Local LLM — 2026 Edition 9 March 2026

  HackerNoon publishes an updated comprehensive guide for running local LLMs, covering current best practices and tooling in 2026. The guide serves as a practical reference for practitioners setting up self-hosted inference systems.

• Nota AI to Showcase End-to-End On-Device AI Optimization at Embedded World 2026 9 March 2026

  Nota AI will demonstrate complete on-device AI solutions from edge optimization to industrial deployment at Embedded World 2026. The showcase highlights production-ready approaches for deploying optimized AI across constrained hardware environments.

• Nemotron 9B Powers Large-Scale Local Inference: Patent Classification and Real-Time Applications 9 March 2026

  Practitioners are leveraging Nemotron 9B for production workloads, from classifying 3.5M patents on a single RTX 5090 to powering real-time Minecraft agent control, demonstrating the model's efficiency and practical viability.

• Qwen 3.5 Family Benchmark Comparison Shows Strong Performance Across Smaller Models 9 March 2026

  New benchmarks reveal that Qwen 3.5's 27B, 35B, and 122B variants retain most of the flagship model's performance, while smaller 2B and 0.8B models show steeper degradation on long-context and agent tasks.

• Student Researcher Achieves 42x Model Compression Through Novel Architecture 8 March 2026

  A high school student has developed an architectural approach that reportedly compresses a 17.6 billion parameter model down to 417 million parameters, potentially offering significant implications for edge deployment if the claims hold under peer review.

• Snapdragon Wear Elite Unveiled at MWC 2026, Advancing Wearable AI Inference 8 March 2026

  Qualcomm's Snapdragon Wear Elite processor brings enhanced AI capabilities to wearable devices. The new chip enables lightweight model deployment on smartwatches and fitness trackers.

• Samsung Opens Registration for Vision AI QLED and OLED Television Integration 8 March 2026

  Samsung introduces Vision AI capabilities in its QLED and OLED televisions, bringing on-device AI inference to smart TV hardware. The move demonstrates expanding edge computing adoption in consumer electronics.

• Windows 11 Notepad Gets On-Device AI Text Generation Without Subscription 7 March 2026

  Microsoft is bringing on-device AI text generation capabilities to Windows 11 Notepad, powered by local models that don't require cloud subscriptions. This mainstream OS integration signals growing adoption of edge AI.

• Alibaba Releases Qwen 3.5 AI Model with On-Device AI Support 7 March 2026

  Alibaba has released Qwen 3.5, a new AI model designed with on-device inference capabilities. This release expands the ecosystem of locally-deployable models optimized for edge devices and self-hosted environments.

• Building PyTorch-Native Support for IBM Spyre Accelerator 6 March 2026

  IBM Research has developed native PyTorch support for the IBM Spyre Accelerator, enabling optimised local inference on specialised hardware.

• The Emerging Role of SRAM-Centric Chips in AI Inference 6 March 2026

  Hardware architectures optimized around SRAM are reshaping AI inference capabilities for edge and local deployments. This emerging trend addresses critical bottlenecks in memory bandwidth and latency for on-device LLM execution.

• Unity Showcases Manufacturing AI Workflow at Smart Factory Expo 5 March 2026

  Unity demonstrated AI-powered manufacturing workflows at Smart Factory Expo, highlighting edge-based inference applications in industrial settings where latency, reliability, and privacy are critical requirements.

• Kakao Launches Kanana AI for On-Device Schedule and Recommendation Management 5 March 2026

  Kakao introduced Kanana, an on-device AI assistant integrated into KakaoTalk that proactively manages user schedules and provides recommendations, demonstrating practical deployment of local intelligence in consumer messaging platforms.

• Qualcomm Snapdragon Wear Elite Brings On-Device AI to Smartwatches 4 March 2026

  Qualcomm's new Snapdragon Wear Elite chip integrates on-device AI capabilities optimized for wearable devices, extending local inference to ultra-constrained environments. The platform enables efficient model execution on smartwatches without relying on smartphone or cloud connectivity.

• Qwen 3.5-27B Q4 Quantization Comparison and Analysis 4 March 2026

  Community-driven quantization sweep compares multiple GGUF quantization approaches for Qwen 3.5-27B, providing data-driven guidance for selecting optimal quantization formats.

• On-Device AI Laptop Lineups Become Standard Across Major Manufacturers 4 March 2026

  Major laptop manufacturers are releasing new product lines with dedicated on-device AI capabilities, signaling a shift from cloud-dependent computing toward local model execution. The trend reflects growing demand from users and enterprises seeking privacy, latency, and offline-capable AI features.

• Alibaba's Qwen 3.5 Small Model Runs Directly on iPhone 17 3 March 2026

  Alibaba releases Qwen 3.5, a lightweight AI model optimized for on-device inference on Apple's iPhone 17. This breakthrough demonstrates practical edge deployment of capable language models on consumer mobile hardware.

• Qualcomm Snapdragon Wear Elite: 2B Parameter NPU for Personal AI Wearables 3 March 2026

  Qualcomm unveils Snapdragon Wear Elite with a dedicated 2 billion-parameter NPU designed for AI inference on smartwatches and wearables. The platform enables always-on personal AI assistants with 30% improved battery efficiency.

• Apple M4 iPad Air Targets AI Users with Double M1 Speed Performance 3 March 2026

  Apple introduces the M4 chip in iPad Air at $599, doubling M1 performance and enabling sophisticated on-device AI inference. The affordable entry point democratizes local LLM deployment on Apple hardware.

• Change Intent Records: The Missing Artifact in AI-Assisted Development 2 March 2026

  An exploration of how explicitly recording developer intent during AI-assisted coding can improve local model fine-tuning and create better training signals for specialized inference models.

• Critical: Qwen 3.5 Requires BF16 KV Cache, Not FP16 for Accurate Inference 2 March 2026

  Community member Daniel Han alerts users that Qwen 3.5 models require bfloat16 KV cache precision instead of the default float16, with perplexity measurements demonstrating the accuracy impact when using incorrect cache formats.

• Jan Releases Code-Tuned 4B Model for Efficient Local Code Generation and Development Tasks 2 March 2026

  The Jan team open-sources Jan-Code-4B, a specialized 4-billion parameter model fine-tuned for code generation, refactoring, debugging, and test writing while optimizing for local deployment and efficiency.

• AI-Native Store Research 1 March 2026

  An exploration of how AI is being integrated into retail environments, including potential applications of local LLM deployment for edge-based customer interaction and inventory management systems.

• Switch Qwen 3.5 Thinking Mode On/Off Without Model Reload Using setParamsByID 1 March 2026

  Unsloth and Qwen community members have discovered how to toggle thinking vs. instruct mode on Qwen 3.5 without reloading the model, enabling dynamic workflow switching and reducing inference latency.

• Qwen 3.5-35B-A3B Emerges as Efficient Daily Driver, Replacing 120B Models 1 March 2026

  Qwen 3.5-35B-A3B is delivering exceptional performance at one-third the size of previous daily drivers, offering significant efficiency gains for local deployment without sacrificing capability.

• Galaxy S26 Debuts AI-Powered Scam Detection in Bold Security Push 28 February 2026

  Samsung's Galaxy S26 implements on-device AI models for real-time scam detection, demonstrating practical deployment of edge inference for security-critical mobile applications.

• Qwen3.5-35B Successfully Runs on Raspberry Pi 5 at 3+ Tokens/Second 28 February 2026

  Demonstration of Qwen3.5-35B inference on Raspberry Pi 5 (16GB and 8GB variants) achieving over 3 tokens/second, proving high-capacity models viable on edge devices.

• Meta Reveals AI-Packed Smartwatch In 2026 – Why Wearables Shift Now 28 February 2026

  Meta's 2026 smartwatch announcement signals the industry's push toward on-device AI in wearable devices, creating new hardware constraints and opportunities for edge model optimization.

• Arduino, Qualcomm Bring On-Device AI and Robotics Learning to Indian School Systems 28 February 2026

  Arduino and Qualcomm partner to integrate on-device AI and robotics education into Indian schools, democratizing access to edge ML training and embedded systems development.

• Unsloth Dynamic 2.0 GGUFs 28 February 2026

  Unsloth releases Dynamic 2.0 GGUF format models, advancing quantized model optimization for local inference with improved efficiency and compatibility across edge devices.

• Qwen 3.5-27B Demonstrates Exceptional Performance with Thoughtful Prompt Engineering 28 February 2026

  Users report that Qwen 3.5-27B significantly exceeds expected performance for its size when paired with effective prompting strategies, suggesting prompt engineering can bridge the capability gap between model sizes.

• LLmFit: Terminal Tool for Right-Sizing LLM Models to Your Hardware 28 February 2026

  LLmFit is a new command-line tool that automatically detects system hardware specifications and recommends the optimal LLM from a database of 497 models across 133 providers, scoring candidates on quality, speed, fit, and cost.

• Accuracy vs. Speed in Local LLMs: Finding Your Sweet Spot 28 February 2026

  A practical guide exploring the trade-offs between model accuracy and inference speed when deploying LLMs locally, helping practitioners optimize for their specific use cases and hardware constraints.

• Arduino, Qualcomm Bring On-Device AI and Robotics Learning to Indian School Systems 27 February 2026

  Initiative bringing practical on-device AI and robotics education to schools, demonstrating accessible pathways for learning local model deployment on edge hardware.

• Arduino and Qualcomm Bring On-Device AI Learning to Indian Schools 27 February 2026

  Arduino and Qualcomm partner to introduce on-device AI and robotics education in Indian schools, democratizing access to edge AI development skills and hardware platforms.

• Show HN: Caret – Tab to Complete at Any App on Your Mac 27 February 2026

  A new macOS application brings local LLM-powered code completion to any application through a tab-triggered interface, demonstrating practical on-device inference for productivity tools.

• Agent System – 7 specialized AI agents that plan, build, verify, and ship code 26 February 2026

  A new multi-agent system coordinates seven specialized agents to handle planning, development, verification, and deployment of code. This demonstrates practical frameworks for orchestrating local LLMs in complex workflows.

• Apple: Python bindings for access to the on-device Apple Intelligence model 26 February 2026

  Apple releases official Python bindings for accessing its on-device Apple Intelligence model, enabling developers to integrate local inference capabilities directly into applications.

• Running LLMs on Raspberry Pi and Edge Devices: A Practical Guide 26 February 2026

  A practical guide for deploying language models on resource-constrained edge devices like Raspberry Pi, including optimization techniques and real-world deployment patterns. Critical for understanding the limits and possibilities of truly local inference.

• Qwen3.5 122B Achieves 25 tok/s on 72GB VRAM Setup 26 February 2026

  Users report exceptional performance running Qwen3.5 122B across three 3090s with 72GB total VRAM, reaching 25 tokens/second with full GPU loading. The model demonstrates strong inference speed and practical viability for enthusiasts with mid-range hardware stacks.

• DeepSeek Releases DualPath: Addressing Storage Bandwidth Bottlenecks in Agentic Inference 26 February 2026

  A new paper from DeepSeek, Peking University, and Tsinghua University presents DualPath, a technique for breaking storage bandwidth limitations in agent-based LLM inference. The research tackles a fundamental performance constraint affecting local deployment at scale.

• What Breaks When AI Agent Frameworks Are Forced Into <1MB RAM and Sub-ms Startup 25 February 2026

  A deep dive into the fundamental constraints and trade-offs when deploying AI agent frameworks on severely resource-limited devices, exploring what architectural patterns fail and what succeeds at the edge.

• Show HN: 100% LLM Accuracy–No Fine-Tuning, JSON Only 25 February 2026

  A technique for achieving perfect LLM accuracy on structured outputs using JSON schema constraints rather than model fine-tuning, reducing computational overhead for local deployments.

• Enhanced Interface Speed Enables High-Performance On-Device AI Features in Smartphones 24 February 2026

  New interface technologies are delivering significant performance improvements for on-device AI inference on mobile devices, enabling faster and more efficient local LLM execution on smartphones.

• Show HN: Dypai – Build Backends from Your IDE Using AI and MCP 24 February 2026

  Dypai enables developers to build backend infrastructure using AI agents through Model Context Protocol integration, streamlining deployment workflows for local LLM applications. This tooling advance simplifies the infrastructure layer for self-hosted AI deployments.

• Breaking the Speed Limit: Strategies for 17k Tokens/Sec Local Inference 23 February 2026

  New techniques and optimisations enable local LLM inference to achieve 17,000 tokens per second, pushing the boundaries of what's possible on consumer hardware. This breakthrough demonstrates practical strategies for maximising throughput in edge deployments.

• Open-Source Framework Achieves Gemini 3 Deep Think Level Performance Through Local Model Scaffolding 23 February 2026

  A new open-source framework enables local models to achieve Gemini 3 Deep Think and GPT-5.2 Pro-level performance through intelligent model scaffolding and composition techniques.

• Elastic Introduces Best-in-Class Embedding Models for High Performance Semantic Search 23 February 2026

  Elastic releases optimized embedding models designed for local deployment and semantic search applications. These models enable efficient vector search on-device without external API dependencies.

• How Slow Local LLMs Are on My Framework 13 AMD Strix Point 22 February 2026

  A detailed performance analysis of running local LLMs on the Framework 13 laptop with AMD Strix Point processor, revealing real-world inference speed benchmarks and practical considerations for edge deployment on modern mobile hardware.

• AI PCs Explained: 7 Critical Truths About NPUs and Privacy 22 February 2026

  A deep dive into NPU-equipped AI PCs and the privacy implications of on-device inference, clarifying misconceptions about local AI processing capabilities.

• I Thought I Needed a GPU to Run AI Until I Learned About These Models 21 February 2026

  A practical guide demonstrating that modern optimized models and inference engines enable effective LLM deployment on CPU-only hardware, removing a major perceived barrier to local AI.

• [Release] Ouro-2.6B-Thinking: ByteDance's Recurrent Model Now Runnable Locally 21 February 2026

  ByteDance's novel recurrent Universal Transformer architecture (Ouro-2.6B-Thinking) is now functional for local inference after fixes for transformers 4.55, enabling access to a unique thinking-focused model on consumer hardware.

• I Stopped Paying for ChatGPT and Built a Private AI Setup That Anyone Can Run 20 February 2026

  MakeUseOf features a detailed account of building a self-hosted LLM alternative to ChatGPT, demonstrating accessible methods for local inference that reduce dependency on cloud APIs.

• VaultAI – 42 AI Models on a Portable SSD, Works Offline for $399 20 February 2026

  VaultAI packages 42 AI models on a portable SSD enabling complete offline inference without cloud dependencies. This represents a practical solution for on-device deployment with minimal hardware requirements.

• The Path to Ubiquitous AI (17k tokens/sec) 20 February 2026

  A technical analysis of achieving 17,000 tokens per second inference throughput, demonstrating the performance milestones required for truly practical local LLM deployment at scale.

• Mirai Secures $10M to Optimize On-Device AI Amid Cloud Cost Surge 20 February 2026

  Mirai, founded by creators of Reface and Prisma, raises $10M Series A funding to advance on-device AI inference optimization, addressing the market shift toward edge computing and away from cloud-dependent models.

• Sarvam Brings AI to Feature Phones, Cars, and Smart Glasses 19 February 2026

  Sarvam AI demonstrates practical on-device AI deployment on ultra-resource-constrained devices, from feature phones to automotive and wearable platforms.

• Running Local LLMs and VLMs on Arduino UNO Q with yzma 19 February 2026

  A new guide demonstrates running local LLMs and vision language models on the Arduino UNO Q microcontroller using yzma. This pushes edge inference to the extreme lower end of hardware constraints.

• AI Integration in Sublime Text: Practical Local LLM Editor Enhancement 19 February 2026

  A developer shares practical techniques for integrating local AI models directly into Sublime Text for code completion and assistance. This shows how local LLMs are being embedded into developer workflows.

• Enhanced Quantization Visualization Methods for Understanding LLM Compression Trade-offs 19 February 2026

  Community members have developed improved visualization techniques for quantization methods, providing clearer insights into how different compression strategies affect model performance and inference characteristics.

• Tailscale Releases New Tool to Prevent Sensitive Data Leakage to Cloud AI Services 18 February 2026

  Tailscale has developed a tool designed to ensure organizations can keep sensitive data local while preventing accidental exposure to cloud AI APIs, reinforcing the security case for local inference.

• GLM-5 Technical Report: DSA Innovation Reduces Training and Inference Costs 18 February 2026

  Alibaba releases GLM-5 technical report detailing key innovations including DSA adoption that significantly reduces training and inference costs while maintaining long-context fidelity.

• Qwen3-Next 80B MoE Achieves 39 Tokens/Second on RTX 5070/5060 Ti Dual-GPU Setup 17 February 2026

  A community member has optimised Qwen3-Next 80B mixture-of-experts to run at 39 tokens/second on dual RTX 50-series GPUs with 32GB total VRAM, sharing previously undiscovered configuration solutions for consumer-grade hardware.

• Alibaba Unveils Major AI Model Upgrade Ahead of DeepSeek Release 16 February 2026

  Alibaba has announced a significant upgrade to its AI models, intensifying competition in the open-source and local deployment space as DeepSeek prepares its latest release.

• NVIDIA's Dynamic Memory Sparsification Cuts LLM Inference Costs by 8x 14 February 2026

  NVIDIA introduces Dynamic Memory Sparsification technique that reduces LLM reasoning costs by 8x through intelligent KV cache management without accuracy loss.

• MiniMax-M2.5 230B MoE Model Released with GGUF Support for Local Deployment 14 February 2026

  MiniMax-M2.5, a 230B parameter mixture-of-experts model, is now available in GGUF format for local deployment with impressive performance benchmarks on consumer hardware.

• Optimal llama.cpp Settings Found for Qwen3 Coder Next Loop Issues 13 February 2026

  Community discovers optimal llama.cpp configuration to fix repetitive loop problems in Qwen3-Coder-Next models, improving practical deployment reliability.

• The Future of AI Slop Is Constraints - Implications for Local Models 13 February 2026

  Analysis of how constraints and optimization techniques are becoming crucial for effective AI deployment, particularly relevant for resource-limited local inference.

• MiniMax M2.5: 230B Parameter MoE Model Coming to HuggingFace 13 February 2026

  MiniMax officially confirms open-source release of M2.5, a 230B parameter MoE model with only 10B active parameters, showing impressive SWE-Bench performance at 80.2%.

• Ming-flash-omni-2.0: 100B MoE Omni-Modal Model Released 13 February 2026

  Ant Group releases Ming-flash-omni-2.0, a 100B MoE model with 6B active parameters supporting unified speech, SFX, music generation alongside image, text, and video processing.

• Energy-Based Models Compared Against Frontier AI for Sudoku Solving 11 February 2026

  New analysis compares specialized energy-based models with large frontier AI systems for Sudoku solving, exploring efficiency advantages of task-specific local models.

• NAS System Achieves 18 tok/s with 80B LLM Using Only Integrated Graphics 11 February 2026

  A community member successfully runs an 80B parameter language model on a NAS system's integrated GPU at 18 tokens per second, demonstrating efficient local inference without discrete graphics cards.