Ollama v0.32.6 リリースOllama Releases v0.32.6

Ollama has published version 0.32.6, a point release that focuses on two areas: faster inference for the Qwen3.5 model on Apple hardware and closer conformance to OpenAI's streaming API format. For developers running large language models locally, both changes address practical friction points—performance on Apple Silicon and drop-in compatibility with existing OpenAI client code.

The headline change is that Qwen3.5 now runs faster on Apple GPUs. According to the release notes, the MLX engine automatically uses the model's MTP head for speculative decoding. MLX is Apple's open-source array framework built specifically for Apple Silicon, and Ollama's MLX engine path lets it take advantage of the unified memory and GPU on M-series chips rather than relying solely on more general backends. Enabling this behavior automatically means users should benefit without needing to configure a separate draft model or tune additional flags.

Speculative decoding is a technique for accelerating text generation without changing the output distribution. In the conventional form, a small, fast "draft" model proposes several tokens ahead, and the larger target model verifies them in a single forward pass, accepting the ones that match and discarding the rest. Because verification is cheaper than sequential generation, the approach can raise throughput when the draft's guesses are frequently correct. The MTP, or multi-token prediction, head is a component trained into some newer models—Qwen3.5 among them—that predicts more than one future token at a time. Using that built-in head as the draft source avoids the need to load and manage a separate small model, which simplifies deployment and reduces memory overhead.

The second change concerns the OpenAI-compatible endpoint. Ollama exposes a /v1/chat/completions route that mirrors OpenAI's API so that tools written for OpenAI can point at a local Ollama server with minimal modification. In this release, the streaming responses from that endpoint were adjusted to match OpenAI's wire format more precisely. Specifically, the role field now appears only on the first chunk, the finish_reason is delivered on its own chunk, and token usage information is sent in a separate chunk. These may sound like small details, but strict client libraries and downstream parsers often expect exactly this sequence; mismatches can cause errors or dropped metadata in applications that were built and tested against OpenAI's own service.

Ollama is a widely used tool for running open-weight language models on local machines, packaging model weights, a runtime, and a REST API behind a single command-line interface. It competes with and complements other local-inference options such as llama.cpp, LM Studio, and server-oriented stacks like vLLM. A recurring theme across these projects is the effort to be a drop-in replacement for hosted APIs, which is why OpenAI wire-format compatibility carries weight: it lets developers reuse SDKs, agents, and integrations without rewriting client code.

Apple Silicon support has become a significant front in local inference because the unified-memory architecture of M-series chips allows relatively large models to run on consumer laptops and desktops. By routing Qwen3.5 through the MLX engine and leaning on the model's own MTP head, Ollama appears to be targeting the segment of users who run models directly on Macs and want better token rates without extra setup.

The Qwen family, developed by Alibaba, has been iterated rapidly and is frequently among the first open-weight releases to adopt features like multi-token prediction. Support for such capabilities in a runtime like Ollama tends to follow model releases, and this update reflects that pattern: the engine is being extended to exploit architectural features the models already ship with.

As with any point release, the practical gains will depend on hardware, model quantization, and workload. Users on non-Apple platforms are unlikely to see the MLX-specific speedups, and the streaming fixes will mainly matter to those consuming the OpenAI-compatible endpoint programmatically. Anyone relying on that endpoint in production may want to verify their client's behavior against the revised chunk sequence. The release is available through Ollama's standard update channels.