Ollama v0.32.6 accelerates Qwen3.5 on Apple GPUs, fixes OpenAI streaming
Qwen3.5 now uses MLX speculative decoding automatically, making Mac inference faster
Ollama, the popular local LLM runtime, released v0.32.6 with a major performance boost for Apple Silicon users. The update enables automatic speculative decoding on the MLX engine by leveraging Qwen3.5's MTP (multi-token prediction) head. This means Qwen3.5 models can now generate tokens faster on Macs without any user-side configuration. The MLX engine and the underlying llama.cpp engine were both updated, reflecting Ollama's continued focus on optimizing inference across hardware backends.
On the API side, Ollama tightened OpenAI compatibility. The /v1/chat/completions streaming endpoint now matches OpenAI's wire format precisely: the role appears only on the first chunk, finish_reason is sent in its own final chunk, and usage data is delivered separately when stream_options.include_usage is enabled. Truncated responses now report finish_reason: "length" instead of "tool_calls", fixing a subtle but important bug for agentic workflows. Additionally, `ollama run kimi-k3` now gracefully falls back to a dedicated kimi-k3:cloud tag for cloud-only models instead of failing. TUI fixes include proper rendering of pipe-delimited prose, Enter accepting highlighted file completions, and lag-free /prompt scrolling. The trade-off: experimental image generation was temporarily removed, so users needing that feature should stay on 0.32.5.
- Qwen3.5 gets automatic speculative decoding via MLX engine's MTP head on Apple GPUs, improving token generation speed.
- OpenAI-compatible streaming now matches wire format: role on first chunk, finish_reason on final chunk, usage in separate chunk.
- Truncated responses now correctly report finish_reason: "length" instead of "tool_calls"; experimental image generation removed.
Why It Matters
Developers running local LLMs on Macs get faster inference and drop-in OpenAI streaming compatibility, simplifying migration and production use.