Ollama changelog

Claude Desktop Developers can now easily configure Claude Desktop to seamlessly work with Ollama as a third-party gateway provider. Improved caching Fixed a hang where agent clients that cancel long prefills Prefill restore points are now t

New desktop onboarding flow on first launch Caches resolved model metadata between requests, cutting time-to-first-token by roughly half (TTFT dropped from ~995 ms to ~524 ms in benchmarks) Fixes a bug where chat and generate could wedge af

llm: transcode WebP images for llama-server renderers/qwen: tolerate non-leading system messages

qwen3.8: support developer instructions

Qwen 3.8 27B This release adds the support of Qwen 3.8 27B. Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. ollama run qwen3.8:27b For Apple Silicon devices, Ollama has in parti

ollama launch dsh now supports DeepSeek Harness, DeepSeek's open-source agent harness ollama launch muse now supports Muse Code, Meta's agentic coding CLI The OpenAI-compatible Responses API now supports web search Muse Glimmer template upd

NVIDIA Nemotron 3.5 Lightning NVIDIA Nemotron 3.5 Lightning is an open 30B mixture-of-experts (MoE) model with 3B active parameters built for that execution layer of always-on agents. It is designed for harnesses like OpenClaw and Hermes Ag

Muse Glimmer Muse Glimmer is now available on all platforms. Muse Glimmer can power coding agent applications such as Claude Code, Codex, Pi and more, as well as long-running personal assistants such as OpenClaw and Hermes. Ollama's MLX eng

Muse Glimmer Note: Muse Glimmer is currently available via initial support via Ollama's MLX engine on Apple Silicon. Additional support and optimizations for Apple Silicon, NVIDIA, AMD, and other platforms will be available in the coming da

Qwen3.5 is faster on Apple GPUs: the MLX engine now uses the model's MTP head for speculative decoding automatically /v1/chat/completions streaming now matches OpenAI's wire format: role only on the first chunk, finish_reason on its own chu

Fixed an MLX Metal bug that could reduce output quality for NVFP4 models, particularly Laguna.

Support Laguna on Apple GPUs via the MLX engine Quantize draft-model output heads at the requested type when creating speculative-decoding drafts. Fixed Qwen3 MoE decoding for differently-quantized experts, plus faster packed gate/up projec

Fixed model downloads that stall before sending data. Improved integrations: restored Claude Code Channels, fixed Anthropic thinking streams, and made Hermes Desktop respect --force-build. Expanded GPU support with CUDA on Windows ARM64, B2

Improved Gemma 4 tool calling and multi-turn reasoning, including more reliable tool-response continuations Fixed a recurrent MLX model cache leak that could increase memory use across requests, and improved cache snapshot performance MLX t

New interactive agent experience: running ollama now launches an agent to help you code and delegate work ❯ ollama Ollama 0.32.0 ▸ Chat, Code, & Work (glm-5.2:cloud) Chat with models, code, search the web, and delegate real work Renamed the

Enabled flash attention on older NVIDIA GPUs (compute capability 6.x) iGPU can now offload vision models with padding to fit available memory Fixed structured output for thinking models when thinking is disabled Hardened GGUF model creation

Faster Gemma 4 on Apple Silicon Gemma 4 is now significantly faster in Ollama on Apple Silicon, generating tokens nearly 90% faster on average across a coding-agent benchmark by leveraging multi-token prediction (MTP). Ollama auto-tunes how

launch: add thinking capability detection to opencode by @hoyyeva in https://github.com/ollama/ollama/pull/15434 launch: auto-install Claude Code by @hoyyeva in https://github.com/ollama/ollama/pull/16802 launch: auto-install opencode when

Command A and North family models now run on Apple Silicon with the MLX engine Updated the underlying llama.cpp engine to build 9672 Fixed build artifacts for MLX

Support for Cohere2Moe architecture Fixed LFM2 parser/render for cases where thinking was not emitted Fixed issue where ollama launch claude and other coding agent or assistant use cases would only output one token Ollama will now return an

Fixed ollama launch selecting the wrong provider in some cases Improved prompt caching by decoupling it from context shift for better KV cache reuse More stable MLX inference with hardened linear and embedding layers MLX runner now creates

Ollama Launch now supports Hermes Desktop, a native desktop interface for the Hermes agent. Run it alongside your Hermes agent to get a visual interface for managing conversations, integrations, and messaging apps. ollama launch hermes-desk

New models Gemma 4 QAT weights: the Gemma 4 family is now optimized with Quantization-Aware Training (QAT) to dramatically reduce memory requirements and maximize on-device performance. Look for the tags ending in -qat: gemma4:e2b-it-qat ge

Fixed the gemma4:12b floating point exception crash on x86, CUDA, Linux, and Windows systems. ollama launch hermes-desktop now launches Hermes Desktop and can skip rebuilding when a packaged desktop app is already installed. ollama launch h

New models Nemotron-3-Ultra: NVIDIA Nemotron 3 Ultra is built for high-throughput reasoning and long-running agent workflows. What's Changed Fixed multimodal models not using GPU on the llama.cpp backend can now use Metal GPU offload on App

New models Gemma 4 12B: high-performance multimodal intelligence that runs directly on laptops, combining efficiency with advanced reasoning. What's Changed Added support for gemma4:12b.

ollama launch now supports Qwen Code and can guide users through installing the Cline CLI when it is missing. ollama launch codex now uses an isolated launch configuration, avoiding conflicts with a user's existing Codex settings. Added lla

Codex App Ollama 0.24 includes support for the Codex App, OpenAI's desktop experience for working on Codex threads in parallel with built-in worktree support and git functionality. ollama launch codex-app Built-in browser Codex can load loc

Ollama 0.30 is now available, with improved compatibility and performance using llama.cpp. This augments the MLX engine on Apple Silicon, bringing support to a wider range of hardware. This release brings support for a wider range of models

ollama launch opencode now supports vision models with image inputs Fixed formatting of Claude tool results when using local image paths

mlx: refined model push behavior by @dhiltgen in https://github.com/ollama/ollama/pull/15431 test: integration test hardening by @dhiltgen in https://github.com/ollama/ollama/pull/13532 app: harden update flows by @dhiltgen in https://githu

ollama launch no longer includes Claude Desktop due to the third-party integration being limited to Anthropic models. Use ollama launch claude-desktop --restore to restore Claude Desktop to its normal state. /api/show responses are now cach

Gemma 4 MTP (Multi-token Processing) for the MLX runner Gemma 4 MTP speculative decoding is now supported on Macs. This can give over a 2x speed increase for the Gemma 4 31B model on coding tasks. ollama run gemma4:31b-coding-mtp-bf16 What'

Claude Desktop Claude Desktop is now supported with Ollama Launch. Claude Cowork and Claude Code are supported within the Claude Desktop App. ollama launch claude-desktop Claude Cowork Claude Code Claude Code on the terminal can still be ac

Updated the Gemma 4 renderer for thinking and tool calling improvements Model recommendations are now updated without updating Ollama Aligned the desktop app's launch page with ollama launch integrations Fixed the Poolside integration title

New models NVIDIA's Nemotron 3 Omni Poolside's first open-weight coding model - Laguna XS.2

api: accept "max" as a think value by @ParthSareen in https://github.com/ollama/ollama/pull/15787 openai: map responses reasoning effort to think by @ParthSareen in https://github.com/ollama/ollama/pull/15789

Improved reliability of the OpenClaw onboarding flow in ollama launch Recommended models in ollama launch now appear in a fixed, canonical order OpenClaw integration now bundles Ollama's web search plugin in OpenClaw

Kimi CLI You can now install and run the Kimi CLI through Ollama. ollama launch kimi --model kimi-k2.6:cloud Kimi CLI with Kimi K2.6 excels at long horizon agentic execution tasks through a multi-agent system. MLX runner adds logprobs suppo

Hermes Agent ollama launch hermes Hermes learns with you, automatically creating skills to better serve your workflows. Great for research and engineering tasks. What's Changed Gemma 4 on MLX. Added support for running Gemma 4 via MLX on Ap

Fix quality of gemma:e2b and gemma:e4b when thinking is disabled ROCm: Update to ROCm 7.2.1 on Linux by @saman-amd in https://github.com/ollama/ollama/pull/15483

Gemma 4 tool calling ability is improved and updated to use Google's latest post-launch fixes Parallel tool calling improved for streaming responses Hermes agent Ollama integration guide is now available Ollama app is updated to fix image a

OpenClaw channel setup with ollama launch What's Changed OpenClaw channel setup: connect WhatsApp, Telegram, Discord, and other messaging channels through ollama launch openclaw Enable flash attention for Gemma 4 on compatible GPUs ollama l

mlx: Improve M5 performance with NAX gemma4: enable flash attention

Gemma 4 Tool Calling improvements Added latest models to Ollama App OpenClaw fixes for launching TUI

app: default app home view to new chat instead of launch by @jmorganca in https://github.com/ollama/ollama/pull/15312

Gemma 4 Effective 2B (E2B) ollama run gemma4:e2b Effective 4B (E4B) ollama run gemma4:e4b 26B (Mixture of Experts model with 4B active parameters) ollama run gemma4:26b 31B (Dense) ollama run gemma4:31b What's Changed docs: update pi docs b

Ollama is now powered by MLX on Apple Silicon in preview Ollama on Apple silicon is now built on top of Apple’s machine learning framework, MLX, to take advantage of its unified memory architecture. https://github.com/user-attachments/asset

ggml: force flash attention off for grok by @rick-github in https://github.com/ollama/ollama/pull/15050 mlx: fix KV cache snapshot memory leak by @jessegross in https://github.com/ollama/ollama/pull/15065 mlxrunner: schedule periodic snapsh

Visual Studio Code Microsoft Visual Studio Code now directly integrates with Ollama via GitHub Copilot. If you have Ollama installed, any local or cloud model from Ollama can be selected for use within visual studio code. What's Changed GLM

Add extra check to ensure npm and git are installed before installing OpenClaw Claude Code will now be faster when run locally, due to preventing cache breakages Fix to correctly support ollama launch openclaw --model <model> Register Ollam

Web Search and Fetch in OpenClaw Ollama now ships with web search and web fetch plugin for OpenClaw. This allows Ollama's models (local or cloud) to search the web for the latest content and news. This also allows OpenClaw with Ollama to be

Ollama 0.18 includes improved performance for OpenClaw and Ollama’s cloud models, including the new Nemotron-3-Super model by NVIDIA designed for high-performance agentic reasoning tasks. Improved OpenClaw performance with Kimi-K2.5 This re

parsers: repair unclosed arg_value tags in GLM tool calls by @BruceMacD in https://github.com/ollama/ollama/pull/14656 Reapply "don't require pulling stubs for cloud models" again by @jmorganca in https://github.com/ollama/ollama/pull/14608

Allow thinking levels such as "medium" to correctly interpreted in Ollama's API for all thinking models Add context length to support compaction when using ollama launch

Fixed issue where GLM-OCR would not work due to incorrect prompt rendering Fixed tool calling parsing and rendering for Qwen 3.5 models

New models Qwen3.5: the small Qwen 3.5 model series is now available in 0.8B, 2B, 4B and 9B parameter sizes. What's Changed Fixed crash in Qwen 3.5 models when split over GPU & CPU Fixed issue where Qwen 3.5 models would repeat themselves d

New models Qwen 3.5: a family of open-source multimodal models that delivers exceptional utility and performance. LFM 2: LFM2 is a family of hybrid models designed for on-device deployment. LFM2-24B-A2B is the largest model in the family, s

Fixed issue where tool calls in the Qwen 3 and Qwen 3.5 model families would not be parsed correctly if emitted during thinking

Fixed issue where Ollama's app on Windows would crash when a new update has been downloaded