Ollama Release Notes: What's New in Each Version

Q: How do I check which version of Ollama I have?

Run ollama -v in a terminal. A running server also reports its version at http://localhost:11434/api/version, which is useful when the command-line client and the server differ.

Q: How do I install a specific Ollama version or a pre-release?

On Linux, set OLLAMA_VERSION when you run the install script, for example curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.40.11 sh. Pre-releases (tags ending in -rc) are test builds published on GitHub before the stable release; this page summarizes a release once it ships as stable.

An unofficial summary by Ollama Herd, not affiliated with Ollama. The official release notes are on GitHub.

Ollama ships often: 33 stable releases in the last three months. This page puts each one in a few plain sentences, newest first, so you can see what changed without reading every pull request.

Covers every stable Ollama release from July 27, 2026 to October 24, 2026 (33 releases), newest first, plus a summary of each older release line.

What's new in the latest Ollama (v0.40.11)

Ollama v0.40.11 October 24, 2026

Adds speech generation on macOS: ollama run can read text or a chat model's reply aloud while the audio is being generated.
New OpenAI-compatible /v1/audio/speech endpoint lets apps stream the generated audio.
Adds the d3 decision models (2B, 4B, 9B and 27B) from vLLM Semantic Router, which classify, answer yes or no, or score options over text, state or images.

Official v0.40.11 notes on GitHub

Every release from the last 90 days

Ollama v0.40.10 October 23, 2026

Adds nemotron-3.5-bolt:e4b, NVIDIA's smaller Gemma 4-based vision model with tool use and a 128K-token context window.
The model ships with a DFlash draft for speculative decoding.

Official v0.40.10 notes on GitHub

Ollama v0.40.9 October 22, 2026

flint now runs on Linux and Windows, not only macOS.
Streaming OpenAI Responses and Anthropic Messages requests now return an error when generation fails, instead of stopping silently after partial output.
ollama launch now passes cloud models' context length and capabilities to coding tools, fixing wrong context and compaction settings.
Web search through the Anthropic-compatible API now respects the model's thinking support on follow-up requests.

Official v0.40.9 notes on GitHub

Ollama v0.40.8 October 21, 2026

Adds flint, Recursive's Apache 2.0 coding model: a 40B mixture-of-experts with about 2.5B active parameters and a 262K-token context window.
It targets agentic work across a repository, with reasoning and multi-turn tool use.
At release it ran on macOS only; Linux and Windows support followed in v0.40.9.

Official v0.40.8 notes on GitHub

Ollama v0.40.7 October 18, 2026

Adds pplx-decider, a 27B decision model from Perplexity built on Qwen3.8-27B.
ollama show and the API (new renderer, parser and template_source fields) now report how a model's prompts are rendered and parsed, and where its chat template comes from.
Fixes several OpenAI-compatible Responses API problems, including switching between thinking and non-thinking models mid-conversation and misordered tool call events in streams.
Requests to remote models are now cancelled on ollama.com when the client disconnects.
Speculative decoding works with more draft models, and ollama launch claude keeps CLAUDE_CODE_AUTO_COMPACT_WINDOW within Claude Code's allowed range.

Official v0.40.7 notes on GitHub

Ollama v0.40.6 October 16, 2026

On Windows, stopping ollama serve no longer leaves orphaned llama-server or MLX runner processes behind, even after a forced kill.
Adds the GGUF Q2_0 tensor and file types, so Q2_0 models can be created and loaded.
Fixes ollama run --ask failing on cloud decision models such as clef:cloud.

Running several machines? Ollama Herd's Windows nodes need this version or newer, together with Herd 0.11.3 or newer. Windows node requirements.

Official v0.40.6 notes on GitHub

Ollama v0.40.5 October 14, 2026

Adds nemotron-3.5-super, NVIDIA's 120B (12B active) hybrid Mamba-Transformer vision-language model with tool calling, optional thinking and up to 1M tokens of context.
Adds nemotron-3.5-bolt, a 74B (9B active) distilled sibling sized for a 24 GB NVIDIA GPU, with a bundled drafter for speculative decoding.
Decision models can run on Ollama's cloud as clef:cloud and clef-flash:cloud without a local download.
Fixes Responses API conversations that include custom tool calls, and explains cloud sign-in, payment and upstream access errors more clearly.

Official v0.40.5 notes on GitHub

Ollama v0.40.4 October 12, 2026

ollama run can query decision models from the terminal with --ask: yes or no questions, labels, or scores against a rubric.
ollama launch codex now lists locally installed models in Codex's model picker (switch with /model).
Adds multimodal embeddings through llama-server and fixes loading multimodal embedding models with small context windows.
The MLX runner gains Kev decision model support on Apple Silicon and more reliable speculative decoding.
ollama launch can now use usage credits without a paid plan.

Official v0.40.4 notes on GitHub

Ollama v0.40.3 October 11, 2026

Pauses the background model upgrades added in v0.40.2, which added per-request overhead for high-volume work such as embeddings.
ollama launch claude and ollama launch codex no longer show warnings about Claude connectors or Codex service tiers.

Official v0.40.3 notes on GitHub

Ollama v0.40.2 October 8, 2026

Models pulled with older versions are upgraded in the background the first time they run, for better performance and compatibility on the llama.cpp engine.
The original copy is kept as a backup so downgrading stays safe, which uses extra disk space; the notes include a jq script to delete the backups.
If you delete the backups and later downgrade below 0.40, you need to re-pull those models.
ollama list no longer shows upgraded models twice, and ollama launch claude now uses the model's full context length.

Official v0.40.2 notes on GitHub

Ollama v0.40.1 October 7, 2026

The local server now forwards the cloud /api/usage and /api/balance endpoints for the signed-in account.
On Windows, model manifests are stored as regular files instead of symlinks, which could be created but then fail to open.
Fixes clef-flash decision requests failing on Windows because of file reads past 2 GiB.
CLI onboarding drops its separate account step; cloud models still ask you to sign in.

Official v0.40.1 notes on GitHub

Ollama v0.40.0 September 25, 2026

On Apple Silicon, models whose architecture the MLX engine supports now run on MLX by default.
MLX coverage at release includes qwen3.8, gemma4, qwen3.6 and qwen3.5, plus the decision models nimble, tev1, clef and clef-flash.
embeddinggemma-2 is the first embedding model supported on MLX.
More models are to be enabled on MLX in later releases.

Running several machines? Macs on different Ollama versions can now run the same model on different engines (MLX or llama.cpp), with different speed and memory use, so keep every Mac on the same release. MLX vs Ollama on a Mac.

Official v0.40.0 notes on GitHub

Ollama v0.35.1 September 29, 2026

Adds Cloudflare's clef (27B) and clef-flash (9B) decision models on /v1/systemone, which also accept images alongside the text state.
Modelfiles gain CAPABILITY declarations so model authors can state what a model supports, kept through import, inheritance and export.
Decision models now report only the decision capability, so clients stop offering them for chat, tools or thinking.
Web search can run up to ten searches per response, up from three.

Official v0.35.1 notes on GitHub

Ollama v0.35.0 September 28, 2026

Adds decision models on a new /v1/systemone endpoint, which return choices, probabilities and scores instead of generated text.
The first models are nimble (Bespoke Labs) and tev1 (Together AI), with three question types: choice, noul and score.
Requests that still send the deprecated typical_p parameter now log a warning instead of failing.
Fixes MLX model downloads that stalled indefinitely and the macOS update indicator not showing at startup.

Official v0.35.0 notes on GitHub

Ollama v0.34.4 September 23, 2026

Structured outputs on thinking models are now applied in one pass, making them faster and more reliable.
Qwen 3.8 processes prompts faster on Apple Silicon, and Gemma 4 there now picks an image resolution per image to keep detail.
Fixes intermittent "model not found" errors with large local libraries and a macOS app hang when checking for ChatGPT or Codex.

Official v0.34.4 notes on GitHub

Ollama v0.34.3 September 19, 2026

ollama show and /api/show now list each model's thinking options and its default, including for cloud models.
Nemotron H vision models now run on Apple Silicon with MLX.
Fixes pulling models from Hugging Face, and the macOS app no longer reopens windows you closed.

Official v0.34.3 notes on GitHub

Ollama v0.34.2 September 15, 2026

Running ollama for the first time starts a setup flow to sign in or continue locally, shared with the desktop app on macOS and Windows.
A new ollama://apps link opens the desktop app's Apps page on macOS and Windows.
Fixes runaway memory growth during long generations with MLX speculative decoding.

Official v0.34.2 notes on GitHub

Ollama v0.34.1 September 14, 2026

Creating MLX models from safetensors with ollama create is no longer experimental; making GGUF models from safetensors now requires llama.cpp's own tools.
/api/tags is much faster on large libraries (about 3.1 s down to 294 ms cold in Ollama's testing) and reports capabilities consistently.
typical_p is deprecated: new models can no longer set it, though existing GGUF models keep it.
Repeat-token detection now waits for 100 repeated tokens, reducing false positives in tasks such as OCR.

Official v0.34.1 notes on GitHub

Ollama v0.34.0 September 5, 2026

Ollama models can now be used inside ChatGPT Desktop, set up from the Ollama app on macOS.
The OpenAI-compatible API adds client tool search and response compaction.
Structured output is faster on Apple Silicon.

Official v0.34.0 notes on GitHub

Ollama v0.33.3 September 2, 2026

gemma4 now accepts image and audio input on the MLX engine.
Responses now report cached prompt tokens.
Default parameters defined in a GGUF file are now respected.

Official v0.33.3 notes on GitHub

Ollama v0.33.2 August 27, 2026

The desktop app follows the system appearance again, which restores dark mode.
On macOS, opening the app now hands off to an instance that is already running instead of starting a second one.
The Claude Desktop proxy no longer interrupts in-progress requests when the model catalog updates.

Official v0.33.2 notes on GitHub

Ollama v0.33.1 August 26, 2026

Adds Qwen3.8 Flash Next support on MLX.
The MLX runner gains structured output support.
The MLX runner no longer hits Metal GPU timeouts when loading models from slow storage.

Official v0.33.1 notes on GitHub

Ollama v0.33.0 August 21, 2026

Claude Desktop can now be configured to use Ollama as a third-party gateway provider.
Cancelled prefills keep their cache restore points, so retries resume where they stopped instead of reprocessing the whole prompt.
Fixes a hang when agent clients cancel long prefills, and disables Claude Code's "tokens left" message, which broke the KV cache on every request.
The DeepSeek Harness launcher falls back to npx when a global npm install fails, including on Windows.

Official v0.33.0 notes on GitHub

Ollama v0.32.15 August 19, 2026

Model metadata is now cached between requests, roughly halving time to first token in Ollama's benchmark (about 995 ms to 524 ms).
The desktop app has a new onboarding flow on first launch.
Fixes chat and generate requests getting stuck after a parser error mid-stream.
Qwen 3.8 now handles system messages that are not at the start of the conversation consistently.

Official v0.32.15 notes on GitHub

Ollama v0.32.14 August 15, 2026

WebP images are now transcoded before they are passed to llama-server.
Qwen renderers now accept system messages that are not at the start of the conversation.

Official v0.32.14 notes on GitHub

Ollama v0.32.13 August 14, 2026

Single-change release: Qwen 3.8 now supports developer instructions.

Official v0.32.13 notes on GitHub

Ollama v0.32.12 August 14, 2026

Adds qwen3.8:27b, aimed at coding, research and long-running agent tasks.
An Apple Silicon variant, qwen3.8:27b-mlx, is tuned for coding agents and repeated tasks.

Official v0.32.12 notes on GitHub

Ollama v0.32.11 August 14, 2026

ollama launch dsh sets up DeepSeek Harness, DeepSeek's open-source agent harness.
ollama launch muse sets up Muse Code, Meta's agentic coding CLI.
The OpenAI-compatible Responses API now supports web search.
Updates the prompt template for Muse Glimmer.

Official v0.32.11 notes on GitHub

Ollama v0.32.9 August 11, 2026

Adds nemotron-3.5-lightning, NVIDIA's open 30B mixture-of-experts model (3B active) built for always-on agents such as OpenClaw and Hermes Agent.
Adds the Nemotron 3 architecture and fixes an edge case in Muse Glimmer's function-call parsing.

Official v0.32.9 notes on GitHub

Ollama v0.32.8 August 10, 2026

muse-glimmer now runs on NVIDIA, AMD and other platforms, not only Apple Silicon.
Use it with coding agents and assistants through ollama launch claude, pi, openclaw or hermes with --model muse-glimmer.

Official v0.32.8 notes on GitHub

Ollama v0.32.7 August 10, 2026

Adds Meta's muse-glimmer, a 30B multimodal model built for local agent workloads.
At release it ran only on Apple Silicon through the MLX engine (muse-glimmer:30b-mlx), with DFlash speculative decoding and image input.
Works with ollama launch for Claude Code, Pi, OpenClaw and Hermes.

Official v0.32.7 notes on GitHub

Ollama v0.32.6 August 4, 2026

Experimental image generation has been removed for now; stay on v0.32.5 if you rely on it.
/v1/chat/completions streaming now follows OpenAI's chunk layout: role in the first chunk only, finish_reason in its own chunk, and usage separately with stream_options.include_usage.
Truncated OpenAI-compatible responses now report finish_reason: "length" instead of "tool_calls".
Qwen3.5 is faster on Apple GPUs because MLX now uses the model's MTP head for speculative decoding automatically.
ollama run on a cloud-only model with no default tag, such as kimi-k3, now offers the :cloud tag instead of failing.

Official v0.32.6 notes on GitHub

Ollama v0.32.5 July 27, 2026

Single-fix release: corrects an MLX Metal bug that could lower output quality for NVFP4 models, Laguna in particular.

Official v0.32.5 notes on GitHub

Earlier release lines

One line per minor version. Each links to that line's first release on GitHub.

Line	First release	What changed
0.31	June 30, 2026	Gemma 4 on Apple Silicon became about 90% faster at generating tokens in Ollama's coding-agent benchmark, using multi-token prediction with no configuration. 0.31.2 enabled flash attention on older NVIDIA GPUs (compute capability 6.x) and made `ollama launch` turn off Claude Code telemetry by default.
0.30	May 13, 2026	Began running models through llama.cpp alongside the MLX engine on Apple Silicon, widening hardware and model support (including GGUF models from Hugging Face) and speeding up NVIDIA GPUs. Patches added Gemma 4 12B and QAT weights, Nemotron 3 Ultra, and `ollama launch` support for Qwen Code, Oh My Pi and Hermes Desktop.
0.24	May 14, 2026	Added `ollama launch codex-app` for OpenAI's Codex desktop app, with `--restore` to undo the change. Also reworked the MLX sampler for better output quality on Apple Silicon.
0.23	May 3, 2026	Added `ollama launch claude-desktop` (withdrawn in 0.23.2, since Claude Desktop's third-party mode only works with Anthropic models) and Gemma 4 multi-token prediction for faster decoding on Macs. A patch cached `/api/show` responses, making integrations load faster.
0.22	April 28, 2026	Added NVIDIA's Nemotron 3 Omni and Poolside's Laguna XS.2 coding model. A patch let model recommendations update without upgrading Ollama and improved Gemma 4 thinking and tool calling.
0.21	April 16, 2026	Added Hermes Agent, GitHub Copilot CLI and Kimi CLI to `ollama launch`, and Gemma 4 on MLX. Patches brought logprobs and faster sampling to the MLX runner.
0.20	April 2, 2026	Added Google's Gemma 4 in E2B, E4B, 26B mixture-of-experts and 31B sizes, with patches improving its tool calling and enabling flash attention for it. Patches also added OpenClaw messaging-channel setup through `ollama launch` and moved Linux to ROCm 7.2.1.
0.19	March 27, 2026	Ollama on Apple Silicon began running on Apple's MLX framework, in preview. The release also improved KV cache reuse through the Anthropic-compatible API.
0.18	March 14, 2026	Cloud models no longer need `ollama pull` (a `:cloud` tag connects directly), and `ollama launch` gained `--yes` for scripts and CI. It shipped ROCm 7, which requires updated AMD drivers; patches added web search for OpenClaw and VS Code integration through GitHub Copilot.
0.17	February 21, 2026	`ollama launch openclaw` began installing and configuring OpenClaw automatically, and the desktop apps started choosing a default context length from available VRAM. Patches added Qwen 3.5 and LFM2, with several fixes to Qwen 3.5 tool calling and MLX memory use.
0.16	February 12, 2026	Added an interactive `ollama` command for launching apps, `ollama launch pi`, and the GLM-5 and MiniMax-M2.5 models. Patches added `ollama launch cline` and `OLLAMA_NO_CLOUD=1` (also an app setting) to turn cloud models off.
0.15	January 21, 2026	Introduced `ollama launch`, which sets up Claude Code, Codex, OpenCode and Droid to use Ollama models without manual configuration. Patches added OpenClaw (first named `clawdbot`), passthrough arguments, subagents, and default context lengths based on VRAM.
0.14	January 10, 2026	Added Anthropic API compatibility (`/v1/messages`), an experimental agent CLI behind `ollama run --experimental`, and the `REQUIRES` Modelfile command. Experimental image generation arrived for macOS and Linux (CUDA), starting with Z-Image Turbo.
0.13	November 19, 2025	Added DeepSeek-OCR, a `bench` tool in the repository for measuring model performance, and better GPU and container memory detection. Patches added Ministral 3, Mistral Large 3, Qwen3-Next and Nemotron 3 Nano, and turned flash attention on by default.
0.12	September 18, 2025	Introduced cloud models (preview), a web search API, and log probabilities in both the native and OpenAI-compatible APIs. Patches added Qwen3-VL, opt-in Vulkan support (`OLLAMA_VULKAN=1`) and embedding models in `ollama run`, and dropped macOS 12 and 13.
0.11	August 5, 2025	Added OpenAI's `gpt-oss` 20B and 120B models with native MXFP4 support, then DeepSeek-V3.1 with hybrid thinking and a preview of Turbo cloud hardware. Patches brought new GPU memory estimates (on by default from 0.11.11), `embeddinggemma` and CUDA 13 support.
0.10	July 18, 2025	Introduced Ollama's new desktop app for macOS and Windows, and changed the default for parallel requests (`OLLAMA_NUM_PARALLEL`) to 1. `ollama ps` also began showing each loaded model's context length.

Releases before 0.10 are on Ollama's GitHub releases page.

Frequently asked questions

How do I update Ollama?

On macOS and Windows, Ollama downloads updates on its own: click the menu bar or taskbar icon and choose Restart to update. On Linux, re-run the install script: curl -fsSL https://ollama.com/install.sh | sh. You can also download the latest installer from ollama.com.

How do I check which version of Ollama I have?

Run ollama -v in a terminal. A running server also reports its version at http://localhost:11434/api/version, which is useful when the command-line client and the server differ.

How do I install a specific Ollama version or a pre-release?

On Linux, set OLLAMA_VERSION when you run the install script, for example curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.40.11 sh. Pre-releases (tags ending in -rc) are test builds published on GitHub before the stable release; this page summarizes a release once it ships as stable.

Is this the official Ollama changelog?

No. This is an unofficial summary written by the Ollama Herd project from Ollama's own release notes. Every entry links to the official notes on GitHub, which are the authoritative record.

Why Ollama Herd keeps this page

Ollama Herd routes requests across several machines that each run Ollama, so every Ollama release can change what our users see: a new default, a new engine on Macs, a fix on Windows. We read each release to keep our own docs correct, and publish the summary here. Where a release matters for multi-machine setups, its entry says so. To run Ollama on more than one machine behind a single endpoint, see the quickstart.

Ollama Release Notes