What ollama/ollama shipped
Written by FoxPlug from the repository's own releases; not affiliated with Ollama. Paste your own repository in the box below and FoxPlug writes the same weekly update for your product. This page is an automatic summary of the releases, pull requests and commits in github.com/ollama/ollama; Ollama did not write it and does not use or endorse FoxPlug. Every line links to the change it describes.
Get a weekly update like this for your product, free
Week of September 28, 2026
What shipped
- V0.40.0 released with models on Apple Silicon now running on MLX by default, including support for gemma4, qwen3.6, qwen3.5, and qwen3.8. Release
- MLX tokenizer fixed to match publisher semantics by honoring pretokenizer stages, Unicode boundaries, added-token normalization, and ranked BPE merges. Pull request #18779
- V0.35.1 released adding support for Clef and Clef Flash decision models with multimodal image input capability through /v1/systemone. Release
- MLX decision model improvements reject overflow requests instead of truncating silently and reduce warm latency by ~5ms on M5. Pull request #18776
- CAPABILITY declarations added to Modelfiles and create requests to explicitly define model capabilities across GGUF and safetensors formats. Pull request #18708
- Decision models now report only decision capability in show and list responses to prevent clients from offering them for general chat or tools. Pull request #18737
- Pull command updated to allow release candidates to pull models matching the minimum version by stripping pre-release tags. Pull request #18790
- MLX safetensor model pulls moved out of experimental and fixed stall handling to enable download recovery. Pull request #18625
- MLX runtime bumped to latest version for performance and compatibility improvements. Pull request #18651
Why it matters
V0.40.0 brings significant performance improvements for Apple Silicon users with automatic MLX runtime selection and improved tokenizer accuracy. Decision model support is now more robust with explicit capability declarations and multimodal image inputs, expanding use cases for classification and routing tasks. Infrastructure improvements fix download reliability issues and increase web search capabilities.
Changelog entry
- Models on Apple Silicon now run on MLX by default Release
- MLX tokenizer semantics fixed to match publishers Pull request #18779
- Added Clef and Clef Flash decision model support with image input Release
- Decision model improvements: reject overflow, ~5ms latency reduction Pull request #18776
- CAPABILITY declarations in Modelfiles for explicit capability definition Pull request #18708
- Decision models report only decision capability in API responses Pull request #18737
- Release candidates can now pull models for same version Pull request #18790
- Web search limit increased to 10 per response Pull request #18602
- MLX safetensor pulls moved from experimental, stall recovery fixed Pull request #18625
- SystemOne model support added to MLX Pull request #18701
- MLX runtime updated to latest version Pull request #18651
V0.40.0 is out. Models on Apple Silicon now run on MLX by default. Decision models support images. Tokenizer accuracy improved. Download reliability fixed.
Ollama v0.40.0 is released. On Apple Silicon, supported model architectures automatically run on the MLX runtime for better performance. MLX tokenizer now matches publisher semantics precisely. Decision models gain multimodal image support and clearer capability declarations. Web search integration increased to 10 searches per response, and download stall recovery has been fixed.
Week of September 21, 2026
What shipped
- Models run on MLX on Apple Silicon by default in v0.40.0, with supported architectures automatically using the MLX runtime. Release
- V0.34.4 applies structured outputs on thinking models in a single pass, making them faster and more reliable. Release
- Gemma 4 now selects image resolution dynamically across 70-1120 pixel budgets based on input resolution and aspect ratio. Pull request #18603
- Settings page opens faster by deferring model discovery and showing saved settings first without blocking the UI. Pull request #18598
- Fixed intermittent 'model not found' errors by properly tracking canonicalized model name parts during lookups. Pull request #18438
- The
typical_pparameter now logs a warning instead of hard failing when used in requests. Pull request #18627 - MacOS update icon and menu items now stay in sync at startup to prevent showing conflicting update states. Pull request #18622
- XGrammar updated to 0.2.7 with schema fixes for typed dictionary values and short arrays. Pull request #18615
Changelog entry
- Models run on MLX on Apple Silicon by default (v0.40.0) Release
- Structured outputs on thinking models now apply in single pass (v0.34.4) Release
- Gemma 4 dynamically selects image resolution based on input Pull request #18603
- Settings page defers model discovery for faster loading Pull request #18598
- Fixed intermittent model not found errors in large local libraries Pull request #18438
- Typical_p parameter now logs warning instead of failing Pull request #18627
- MacOS update icon and menu stay synchronized at startup Pull request #18622
- XGrammar updated to 0.2.7 with schema fixes Pull request #18615
Week of September 14, 2026
What shipped
- Server-side MLX imports now support safetensors through the create pipeline with remote upload, draft layers, and transfer limits, while GGUF create is limited to wrapping existing inputs. Pull request #14969
- The API now exposes model thinking levels and defaults through /api/show and ollama show so clients can offer controls matching each model's capabilities. Pull request #18473
- Server allows registry cross-host redirects among allowlisted hosts to improve registry flexibility. Pull request #18533
- Registry and blob transfer redirects now validate target schemes and addresses, re-check DNS on each redirect, and prevent downgrading https to http. Pull request #18512
- The MLX runner now releases freed KV buffers during speculative decode to prevent buffer pool exhaustion. Pull request #18510
- On macOS, closing the window with the red button or Command–W now keeps it closed when switching back with Command–Tab. Pull request #18518
- First-run onboarding is now shown when running ollama without a subcommand, with desktop-aligned copy and optional browser sign-up, shared between CLI and desktop. Pull request #18495
- MLX and MLX-C dependencies updated with a fix for quantized matmul corruption affecting nvfp4 multimodal models. Pull request #18449
Why it matters
The MLX engine graduates from experimental to production status with improved safetensors import support and memory management. Registry security and redirect handling are tightened to prevent downgrade attacks. New API metadata exposes model capabilities so clients can offer appropriate controls.
Changelog entry
- Mlxrunner: Move MLX engine from x/ to top-level production directories Pull request #18489
- Server: Add server-side MLX safetensors imports with remote upload and draft layer support Pull request #14969
- Api: Expose model thinking levels and defaults through /api/show Pull request #18473
- Server: Allow registry cross-host redirects among allowlisted hosts Pull request #18533
- Server: Tighten redirect validation for registry requests with scheme and DNS checks Pull request #18512
- Mlxrunner: Release freed KV buffers during speculative decode to prevent exhaustion Pull request #18510
- App: Keep closed macOS windows from reopening on activation Pull request #18518
- Cli: Add first-run onboarding shared with desktop app Pull request #18495
- Deps: Bump MLX and MLX-C with quantized matmul corruption fix Pull request #18449
- Api: Deprecate
typical_pparameter for new models Pull request #18448 - Tests: Fix metadata deletion test flake Pull request #18493
V0.34.3-rc1: MLX engine moves to production, safetensors imports gain server-side support, registry redirects get stricter security validation, and model thinking levels are exposed through the API.
V0.34.3-rc1 ships significant infrastructure improvements: the MLX engine graduates from experimental status to top-level production use, safetensors imports now work server-side with upload and draft layer support, registry redirects are hardened against downgrade attacks, and the API exposes model thinking levels so clients can offer matching controls. Speculative decode also gains memory efficiency improvements.