NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
Go modules, #1 by repository stars
Latest stable
v0.40.17
Last release today
21 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Some releases are documented
notes for 32 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
3 years old
1371 releases · first in 2024
Drag across the ruler to compare, or tap a release to open it.
Nothing published for this version
Inkling 2 Mini from Thinking Machines is a 26B parameter model with 5.4B active.
Inkling 2 Mini from Thinking Machines is a 26B parameter model with 5.4B active.
Its looped layers give it more computation per token than its size suggests, and it still fits on a Mac at about 20GB. It handles text and images, tool calls, and thinking, and scores 79.2% on SWE-Bench Verified).
ollama run inkling-2-mini
Note: inkling-2-mini is supported on Apple silicon only in this release. PC support will follow in an upcoming release.
ollama launch claude supports Claude Code's deferred tool loading, so MCP tools are loaded when needed instead of all at
once
Ollama on Apple silicon now serves multiple requests at once.
ollama launch opencode sets thinking options from the model's metadata
ollama launch pi refreshes Pi's model list after sign-in and passes along prompt arguments
ollama launch chatgpt --restore asks before switching an active profile and preserves existing settings and chat history;
quitting ChatGPT yourself now lets it exit through its own quit prompt
Running ollama with no arguments in a non-interactive shell prints help instead of erroring
Ctrl+C or Ctrl+G while in a multiline """ block leaves multiline mode instead of swallowing the next line
Changing the context shift setting no longer reloads the model
Vulkan now uses mmap for integrated GPUs, fixing Failed to allocate pinned memory load failures on shared-memory GPUs
On Windows iGPUs, partially offloaded models no longer fail to allocate the KV cache
Tool definitions passed through the API keep JSON schema keywords Ollama doesn't model instead of dropping them
Disabling thinking now maps to the model's lowest supported reasoning level rather than being ignored
Source builds can use prebuilt MLX-C and XGrammar binaries
Full Changelog: v0.40.16...v0.40.17-rc3
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Merge remote-tracking branch 'upstream/main' into release_v0.40.17
Merge remote-tracking branch 'upstream/main' into release_v0.40.17
mlxrunner: add Inkling 2 Mini
mlxrunner: add Inkling 2 Mini
Nothing published for this version
Nothing published for this version
mlxrunner: adapt saved history to memory pressure
Retain Metal prefix history using observed memory headroom and execution
workspace rather than a fixed 8 GiB allowance. Release unused buffers and
cold history when system compression grows, preserving reusable paths and
active requests.
Cover pressure reclamation, snapshot ownership, memory observation, and
cacheless request isolation.
address comments
address comments
Nothing published for this version
mlxrunner: batch speculative verification
Keep drafting active for concurrent generation by sharing target verification while preserving per-request sampling and cache rollback. Learn draft depth for each concurrent workload separately from isolated requests, and preserve probes that short requests never execute.
Materialize owned target features before a reader can pause, and cover acceptance, cancellation, handoffs, mixed decoding and retained-memory boundaries.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
MLX, MLX-C: version bump
MLX, MLX-C: version bump (#19647)
Nothing published for this version
Ripple Preview from Ricursive Intelligence is a 65B dense model built for agentic coding.
Ripple Preview from Ricursive Intelligence is a 65B dense model built for agentic coding.
ollama run ripple-previewollama launch claude always routes Claude Code to the selected modelollama launch explains when a cloud model has been retired/v1/responses respects truncation and no longer silently truncates by default/v1/models returns an empty list instead of null when no models are installedFull Changelog: v0.40.15...v0.40.16
Nothing published for this version
Add the Ripple model, renderer, parser and import registration. Support per-layer NoPE in DFlash drafts. Fuse MLP activation, value blending and atten
Add the Ripple model, renderer, parser and import registration.
Support per-layer NoPE in DFlash drafts.
Fuse MLP activation, value blending and attention gating.
Add reference fixtures and model/runner regression tests.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Xiaomi MiMo's mimo-v2.6 , a 9B model built for code and agent workflows, is now supported in Ollama.
Xiaomi MiMo's mimo-v2.6, a 9B model built for code and agent workflows, is now supported in Ollama.
ollama run mimo-v2.6AutoTrust decision models read text and return choices, yes or no probabilities, and ratings for the questions you supply. The family includes 9b, 35b, and 27b-vl sizes — the 35b and 27b-vl sizes can also read images.
ollama run autotrust-jev "A customer says their card was charged twice." --ask "Is this a billing issue?"Note: autotrust-jev is supported on Apple silicon in this release.
Full Changelog: v0.40.14...v0.40.15
cmd/launch: add Kimi Code CLI support
cmd/launch: add Kimi Code CLI support (#19506)
Drop 002-metal-unary-cast-constexpr.patch: superseded by upstream fix
Drop 002-metal-unary-cast-constexpr.patch: superseded by upstream fix
model: add MiMo V2.6 renderer for Qwen3.5
model: add MiMo V2.6 renderer for Qwen3.5 (#19400)
Nothing published for this version
ollama launch claude now applies the model's output token limit for Claude Code, so cloud models no longer run into the 128k default cap
ollama launch claude now applies the model's output token limit for Claude Code, so cloud models no longer run into the 128k default cap/api/chat are now decoded correctly, including compressed responsestemperature/top_p get the model's own values instead of 1.0, matching /api/chat/v1/responses now preserves readable input content, including prior assistant refusals, instead of flattening itollama launch codex verifies ChatGPT routing before configuring, and reports a clear error when routing is unavailableFull Changelog: v0.40.13...v0.40.14
server: decode cloud web search responses
server: decode cloud web search responses (#19476)
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Liquid AI's LFM3 models, including text-to-speech in lfm3-audio and lfm3-omni , now run on Linux and Windows as well as Apple silicon.
Liquid AI's LFM3 models, including text-to-speech in lfm3-audio and lfm3-omni, now run on Linux and Windows as well as Apple silicon.
Full Changelog: v0.40.12...v0.40.13
This temporary patch works around a regression on older versions of MacOS ggml_metal_library_compile_all: failed to build 'unary' library
This temporary patch works around a regression on older versions of MacOS
ggml_metal_library_compile_all: failed to build 'unary' library
llama.cpp: version bump b11963
llama.cpp: version bump b11963 (#19360)
Nothing published for this version
LFM3 is Liquid AI's new family of small models built for local agents, tool calling, and structured output, on Liquid's new tri-hybrid architecture. T
LFM3 is Liquid AI's new family of small models built for local agents, tool calling, and structured output, on Liquid's new tri-hybrid architecture. The family includes:
ollama run lfm3Liquid AI's d1 models now come in a 2B size, built on LFM3. d1 makes decisions about text and images, and d1-omni adds audio.
ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"Note: lfm3 and d1 2b are supported on Apple hardware in this release. Support for Linux/Windows coming soon.
ollama launch now works with OpenCode v2Full Changelog: v0.40.11...v0.40.12
mlx: add LFM3 and d1 models
mlx: add LFM3 and d1 models (#19333)
Nothing published for this version
Nothing published for this version
llama.cpp: version bump b11901
llama.cpp: version bump b11901 (#19311)
Nothing published for this version
Ollama can now generate speech on macOS. ollama run plays text or a chat model's reply aloud as the audio is generated, and apps can stream audio thro
Ollama can now generate speech on macOS. ollama run plays text or a chat model's reply aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible /v1/audio/speech endpoint.
ollama run phonon "Hello from Ollama."d3 brings vLLM Semantic Router's decision models to Ollama. Give a model text, structured state, or an image, and ask it to classify, answer yes or no, or score options with probabilities. The family includes 2B, 4B, 9B, and 27B sizes.
ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"Full Changelog: v0.40.10...v0.40.11
mlxrunner: import and list Phonon catalog voices
mlxrunner: import and list Phonon catalog voices
docs: document Phonon voices
Nothing published for this version
Nemotron 3.5 Bolt E4B is NVIDIA’s smaller, Gemma 4-based vision-language model for fast local conversations, image understanding, and tool use. It sup
Nemotron 3.5 Bolt E4B is NVIDIA’s smaller, Gemma 4-based vision-language model for fast local conversations, image understanding, and tool use. It supports a 128K-token context window.
The model includes a DFlash draft for speculative decoding.
ollama run nemotron-3.5-bolt:e4b
Full Changelog: v0.40.9...v0.40.10
renderers: add gemma4-nemotron renderer
Add a Gemma 4 renderer variant that renders assistant content before
its tool calls rather than after the tool responses. Models opt in with
RENDERER gemma4-nemotron; gemma4 auto-detection is unchanged.
Adds the corresponding chat template as a test fixture for Jinja parity
checks.
Load a draft-owned output head when its config requires one, while preserving the target-head path used by existing DFlash models.
Expose Gemma 4 hidden-state taps and raw logits for speculative decoding. Report safetensors target precision independently of quantized draft weights.
Flint now runs on Linux and Windows as well as macOS.
Flint now runs on Linux and Windows as well as macOS.
ollama launch now passes cloud models' context lengths and capabilities to coding tools, fixing incorrect context and compaction settings.Full Changelog: v0.40.8...v0.40.9
Replace the prerelease checkpoint path with Gradium's GA format, including default-voice import and speech conditioning. Set the 0.40.9 minimum, updat
Replace the prerelease checkpoint path with Gradium's GA format, including default-voice import and speech conditioning. Set the 0.40.9 minimum, update the speech docs, and test import, inference setup, and rejection of the old checkpoint.
Nothing published for this version
Nothing published for this version
flint: fix replies with thinking disabled
Close the thinking section when think:false so Flint answers directly. Default and think:true prompts are unchanged.
Nothing published for this version
llama.cpp update
llama.cpp update (#19194)
We were still using a pre-release, and the download site is flaky, so lets get current on the latest CUDA Toolkit for WoA support.
We were still using a pre-release, and the download site is flaky, so lets
get current on the latest CUDA Toolkit for WoA support.
Flint brings Recursive's coding model to Ollama. It is designed for the work coding agents do across a repository: understanding unfamiliar code, trac
Flint brings Recursive's coding model to Ollama. It is designed for the work coding agents do across a repository: understanding unfamiliar code, tracing bugs, implementing features, and adding tests. Reasoning, multi-turn tool use, and a 262K-token context window support longer tasks across files and tool interactions.
Flint is an Apache 2.0-licensed, 40B-parameter mixture-of-experts model with approximately 2.5B parameters active per token. In its release evaluations, Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified.
ollama run flintNote: Flint is supported on MacOS. Other GPU support will be enabled in a subsequent release.
Full Changelog: v0.40.7...v0.40.8
Preserve source tool JSON alongside decoded API values so Flint can retain numeric types and schema order in prompts. Existing API marshaling is uncha
Preserve source tool JSON alongside decoded API values so Flint can retain
numeric types and schema order in prompts. Existing API marshaling is unchanged.
Assemble ModelOpt expert banks during safetensors import, and reject direct
Flint NVFP4 creation before conversion. Other models retain their existing
import and quantization paths.
Add optional learned sinks and device causal masks to shared MLX attention.
Existing callers retain the no-sink and host-mask paths; Flint opts into the
new behavior.
One column per month.