Ollama v0.32.6 boosts Qwen3.5 on Apple GPUs with speculative decoding

Ollama shipped v0.32.6-rc0 on August 4, packed with performance and compatibility improvements. The headline change is MLX engine acceleration for Qwen3.5 on Apple GPUs: the engine now automatically uses the model's MTP (multi-token prediction) head for speculative decoding, which speeds up generation by predicting multiple tokens per step. This is a notable win for Mac-based developers running Qwen3.5 locally, as speculative decoding typically reduces latency without sacrificing output quality.

Another major fix aligns Ollama's /v1/chat/completions streaming with OpenAI's wire format. Role is now sent only on the first chunk, finish_reason arrives on its own chunk, and usage data ships in a separate chunk when stream_options.include_usage is enabled. This removes a long-standing pain point for teams using OpenAI SDKs or proxies with Ollama. Additionally, truncated responses now correctly report finish_reason: "length" instead of "tool_calls", preventing client-side errors. Ollama also added a kimi-k3:cloud alias for cloud-only models, fixed several TUI rendering issues, and updated both MLX and llama.cpp engines. However, experimental image generation was temporarily removed—users needing it should stick with v0.32.5.