Performance Boost for Qwen3.5 on Apple GPUs
Ollama’s latest release candidate speeds up Qwen3.5 on Apple hardware by automatically enabling speculative decoding through the MLX engine, which now leverages the model’s MTP head. The MLX and llama.cpp engines have also been updated to their latest versions, ensuring broader compatibility and performance improvements.
Smoothed API Compatibility
The /v1/chat/completions streaming endpoint now fully matches OpenAI’s wire format: the role attribute appears only on the first chunk, finish_reason comes in its own dedicated chunk, and usage statistics are sent in a separate chunk when stream_options.include_usage is set. Additionally, truncated responses now correctly report finish_reason as “length” instead of “tool_calls”, aligning with expected behavior.
Better Model Handling and TUI Enhancements
Running ollama run kimi-k3 no longer fails for cloud-only models that lack a default tag. Instead, it intelligently suggests kimi-k3:cloud as the fallback. The terminal UI also receives several polish fixes: pipe-delimited prose no longer mistakenly renders as a table, pressing Enter now accepts the highlighted @ file completion suggestion, and scrolling through the /prompt history is no longer laggy.
Image Generation Temporarily Paused
Experimental image generation support has been removed in this release. Users who rely on this feature should continue using Ollama 0.32.5 until it returns in a future update.
