Ollama v0.32.10: Faster Prefill Performance on NVFP4 Models with System Config Support

Ollama's latest release (v0.32.10) brings meaningful performance improvements to quantized model inference, particularly for NVFP4-formatted models. The optimization in prefill processing yields 7-8% speed improvements on popular models like Qwen 3.6 and Muse Glimmer, achieved by compiling float-to-int multiplication operations into a single kernel. Additionally, the release defaults repeat penalties to 1.0 (off) unless explicitly configured, bringing better alignment with other inference engines.

A notable quality-of-life improvement is the addition of system-level configuration file support, allowing administrators to set default parameters via INI files with CLI > ENV > model-presets > INI precedence. This makes managing multiple deployments across different machines significantly easier. For practitioners running local inference at scale, these updates provide both measurable speed gains and better operational tooling.

Read the full article on Ollama release.

Source: Ollama release · Relevance: 9/10