what stackpulse tracks
Ollama releases from GitHub
StackPulse watches Ollama release notes and keeps the original source link close to every summary.
Get up and running with large language models locally StackPulse turns upstream changelogs into scannable summaries with risky changes, deprecations, migration notes, and source links.
what stackpulse tracks
StackPulse watches Ollama release notes and keeps the original source link close to every summary.
upgrade risk
Risky changes are separated from normal feature notes so you can scan upgrade impact before changing production dependencies.
migration notes
Migration steps and recommended actions are only shown when the upstream release notes support them.
This prerelease includes improvements to ChatGPT CLI restore and graceful restart, plus updated OpenCode settings using thinking metadata. It also adds concurrent generation support for MLX and uses mmap for Vulkan integrated GPUs, alongside llama.cpp and MLX version bumps.
Users of the ChatGPT CLI, OpenCode, MLX, or Vulkan integrated GPUs may benefit from the listed improvements.
This release adds Ripple Preview, a 65B dense model designed for agentic coding. It also improves memory use for Qwen3.8 Flash Next on Apple silicon, adjusts API and launch behavior, and adds CUDA 13.5 support for Windows on ARM.
Users of Qwen3.8 Flash Next on Apple silicon, Claude Code integrations, the `/v1/responses` and `/v1/models` APIs, and Windows on ARM with CUDA may notice the changes.
Run `ollama run ripple-preview` to try the new model.
This prerelease adds MiMo V2.6 model support. It also changes Linux MLX to load bundled CUDA libraries in-process.
Users running Ollama on Linux with MLX or working with MiMo V2.6 models are affected.
This prerelease adds support for the MiMo V2.6 model. It also updates Linux MLX behavior to load bundled CUDA libraries in-process.
Users running Ollama on Linux with MLX and users seeking to run MiMo V2.6 models are affected.
Review the prerelease if you need MiMo V2.6 support or the Linux MLX library-loading change.
This prerelease adds support for the MiMo V2.6 model. It also updates MLX to load bundled CUDA libraries in-process on Linux.
Users running Ollama on Linux with MLX, and users who want to use MiMo V2.6, are affected.
Ollama adds support for Xiaomi’s MiMo V2.6 model and the AutoTrust decision-model family, including Apple silicon support for autotrust-jev. The release also expands launcher integrations, improves concurrent serving of recurrent and vision GGUF models, and opens the welcome launcher when Ollama is run without arguments.
Ollama users, especially those using supported models, launcher integrations, or recurrent and vision GGUF models, can use the new model support and workflow improvements.
Run `ollama run mimo-v2.6` or `ollama launch` to try the new model and launcher integrations.
This release improves Claude Code and Codex launch behavior, fixes cloud web search response decoding, and makes OpenAI-compatible endpoints honor packaged sampler defaults when clients omit sampling parameters. It also preserves readable input content in `/v1/responses`, adds retry handling for ChatGPT disconnects, and speeds up MLX top-k sampling for large vocabularies.
Ollama users of Claude Code, Codex, cloud web search, OpenAI-compatible endpoints, `/v1/responses`, and MLX may notice more reliable behavior or improved sampling performance.
This prerelease includes fixes for LFM3 GGUF support and llama-server speech. It also updates the bundled llama.cpp version to b11963.
Users of LFM3 GGUF models and llama-server speech functionality are affected by the fixes and support included in this prerelease.
Liquid AI’s LFM3 models are now supported on Linux and Windows, in addition to Apple silicon. This includes text-to-speech models lfm3-audio and lfm3-omni.
Ollama users on Linux and Windows can now run Liquid AI’s LFM3 models, including the audio models.
This prerelease adds DFlash2 support to MLX and improves prompt speed for Gemma, Flint, and Laguna through an MLX/MLX-C version bump. It also preserves small-image detail for compact Gemma 4 vision towers and updates llama.cpp to version b11901.
Users of MLX models, especially Gemma, Flint, and Laguna, may benefit from improved prompt performance and vision handling; llama.cpp users receive the dependency update.
Review and test this prerelease if you use MLX models or llama.cpp.
Adds support for Liquid AI’s LFM3 model family, including text, vision, audio, and multimodal variants, plus 2B d1 decision models. The release also improves Apple silicon prompt processing and image detail, adds MLX DFlash2 speculative decoding support, and includes compatibility and tool-turn fixes.
Users of Liquid AI models and Apple silicon users are affected; LFM3 and d1 2B support is limited to Apple hardware in this release, with Linux and Windows support described as coming soon.
Apple hardware users can try the new models with `ollama run lfm3` or `ollama run d1:2b`.
Ollama can now generate speech on macOS, including playing model replies aloud as they are generated and streaming audio through the OpenAI-compatible speech endpoint. The release also adds the d3 decision-model family, which supports classification, yes/no answers, and probability-based scoring.
macOS users and developers integrating speech or decision-model capabilities with Ollama are affected.
On macOS, try speech generation with `ollama run phonon "Hello from Ollama."` or use the `/v1/audio/speech` endpoint; try a d3 model for decision tasks.
This prerelease adds a Phonon voice catalog and improves Phonon text handling for symbols, numbers, and links. It also includes fixes for Qwen3 Coder function-tag handling and OpenCode context allocation, plus a llama.cpp version update.
Users of Phonon voice features, Qwen3 Coder, OpenCode, and the vLLM-D3 decision path may benefit from the updates.
Review the prerelease if you use the affected integrations or Phonon voice features.
This release adds NVIDIA Nemotron 3.5 Bolt E4B, a Gemma 4-based vision-language model for local conversations, image understanding, and tool use. It supports a 128K-token context window and includes a DFlash draft for speculative decoding.
Ollama users who want to run a local vision-language model can use Nemotron 3.5 Bolt E4B.
Run `ollama run nemotron-3.5-bolt:e4b` to try the model.
Ollama adds speech generation with Phonon on Apple Silicon, including streamed audio through an OpenAI-compatible endpoint. It also introduces d3 decision models and expands Flint support to Linux and Windows, alongside fixes for streaming error reporting and cloud-model configuration in coding tools.
Users of Ollama on Apple Silicon can generate speech, users can try d3 decision models, and Linux and Windows users can run Flint.
Ollama adds Flint, Recursive’s coding model, to its model library. Flint supports reasoning, multi-turn tool use, and a 262K-token context window; this release notes that Flint is supported on macOS, with other GPU support planned for a subsequent release.
Ollama users on macOS can run Flint; users relying on other GPU platforms must wait for support in a subsequent release.
On macOS, run `ollama run flint` to use the model.
This release improves OpenAI-compatible Responses conversation continuity and fixes event ordering for streaming tool calls. It adds model renderer, parser, and template-source reporting, expands speculative decoding support, and updates llama.cpp and MLX.
Ollama users and developers using Responses conversations, streaming tool calls, remote models, model metadata, or speculative decoding are affected by these improvements and fixes.
Upgrade to v0.40.7 to receive the Responses fixes and expanded model support.
This release introduces support for Responses effort on non-thinking models, allows llama-server to infer draft speculative types, and adds PPLX Decider v1.2 support.
Users leveraging non-thinking models, llama-server, or PPLX Decider will be affected.
This release introduces support for q2_0 tensor and file types, launches clef in Ollama's cloud, and implements identifiers for --ask mode question keys. It also includes a fix to prevent orphaned runners on Windows.
Users of Ollama on Windows and those utilizing specific tensor types may be affected by the changes.
Update to v0.40.6 to benefit from new features and fixes.
This release introduces support for NVIDIA Nemotron 3.5 Super, a vision-language model for coding and reasoning, and includes updates to decision models, error handling, and dependencies like llama.cpp and MLX.
Users leveraging vision-language models or cloud-based decision models are most affected by this release.
Update to v0.40.5 to access new features and improvements.