Ollama releases: v0.32.10

TL;DR

Ollama released version v0.32.10, which adjusts the default `repeat_penalty` to 1.0 for better performance and compatibility with other engines, and enhances NVFP4 MLX models' prefill speed by about 7–8%.

Detailed Summary

Ollama released version v0.32.10 of its language model software, addressing the default `repeat_penalty` setting for improved compatibility and decoding speed, while also enhancing prefill efficiency on specific hardware models. This update impacts users by providing more consistent performance across different models and optimizing runtime on NVFP4 MLX devices. The changes are expected to benefit a broader community of developers and end-users relying on Ollama's models for various applications.

Key Points

• • Models without `repeat_penalty` now default to 1.0.
• • Speculative decoding is sped up by setting `repeat_penalty` per model.
• • Faster prefill observed on NVFP4 MLX models with global scale.