Ollama v0.32.6-rc0 adds a useful path for people running Qwen3.5 on Apple GPUs: the MLX engine now uses the model's MTP head automatically for speculative decoding.
The practical idea is straightforward. The MTP head predicts upcoming tokens, and the main model verifies those predictions in batches. Accepted predictions can reduce serial waiting during decode.
Ollama says this makes Qwen3.5 faster. It does not publish a benchmark, test machine, or speedup percentage in the release notes, so claims such as "2x faster" are not supported by this release.
The compatibility change may matter just as much
The same release adjusts streaming behavior for the OpenAI-compatible /v1/chat/completions endpoint.
If your client parses chunks, usage data, or finish_reason, rerun its compatibility tests before rolling the RC into a production workflow. Local inference becomes easier to adopt when its wire behavior is predictable, not only when token generation is faster.
Two reasons to test before upgrading
First, this is a release candidate, not the stable release.
Second, experimental image generation was temporarily removed. Ollama recommends staying on 0.32.5 if that feature is part of your workflow.
A sensible split is:
- Test now if you use an Apple GPU with Qwen3.5 and can isolate RC risk.
- Wait if you depend on image generation or production stability.
- Benchmark prefill and decode separately. MTP can improve decode without changing the part of latency your workload actually feels.
Official Ollama v0.32.6-rc0 release
AI-assisted drafting. Release facts and compatibility boundaries were reviewed before publication.
Top comments (0)