Ollama Adds NVIDIA Nemotron 3.5 Super Vision-Language Model Support

16 October 2026 1 min read

#agents #benchmarks #consumer-gpu #inference-optimization #mixture-of-experts #moe #multimodal-ai #nvidia #ollama #quantization #release #vision-language-model

NVIDIA's Nemotron 3.5 Super represents a significant milestone for local LLM deployment, combining frontier capabilities with practical efficiency. With 120B total parameters but only 12B active per token through its mixture-of-experts design, the model achieves strong performance on coding, reasoning, and tool-use tasks while remaining viable for consumer hardware deployments.

Ollama's integration in v0.40.5 makes this model immediately accessible to local deployment practitioners. The model pairs a hybrid Mamba-2/Transformer backbone with a C-RADIO vision encoder, enabling multimodal reasoning without sacrificing inference speed. This represents the convergence of open-weight frontier models with practical deployment infrastructure.

For practitioners, this release demonstrates that sophisticated agent-capable models can now run locally with proper quantization and inference optimization. The availability through Ollama means users can deploy enterprise-grade vision-language capabilities on-device without cloud dependencies, a crucial development for privacy-conscious and latency-sensitive applications.

Read the full article on Ollama release.

Source: Ollama release · Relevance: 9/10