Ollama 0.40.12~0.40.15: LFM3 Expansion, MLX Speculative Decoding and Sampling Improvements
Posted As of
Ollama released four times between October 27 and November 3. LFM3 expanded to Linux and Windows, DFlash speculative decoding became supported on MLX, and top-k sampling got faster.
What happened
Based on the Ollama GitHub release notes (v0.40.12~v0.40.15). The release list does not show years, so the entries were read as 2026 from the surrounding version numbers.
- v0.40.12 (October 27): Liquid AI's LFM3 family (text, vision, audio, omni) was added. At release it ran only on Apple hardware. Speculative decoding using a DFlash draft model is supported on MLX, and prompt processing on Apple Silicon got faster for several models. (official)
- v0.40.13 (October 28): The LFM3 family (including lfm3-audio and lfm3-omni) also runs on Linux and Windows. (official)
- v0.40.14 (November 3): Top-k sampling on MLX got faster for models with large vocabularies. The OpenAI-compatible endpoint no longer overrides the default sampler values stored in a model. (official)
- v0.40.15 (November 3): Recurrent and vision GGUF models now handle multiple concurrent requests. YaRN long-context scaling was added for Qwen3 and Qwen3.5. MiMo V2.6 (9B) was added, and
ollama launchnow supports Kimi Code CLI. (official)
Not in the release notes: the size of the speed gains and the memory required per model.
Our take
For Mac users, the changes on the MLX path (DFlash speculative decoding and sampling speed) are directly relevant. However, the notes contain no figures, so we cannot say how much faster things got until we measure under the same conditions ourselves.
Sources and related posts
- Ollama GitHub 릴리스 노트 Primary source