What’s new in open models

All LiteLLM Docker images are signed with cosign. Every release is signed with the same key introduced in commit 0112e53.

You can now run Qwen-Image-2.1 locally with Unsloth! This release also includes custom Agent Skills, and easier chat/project management. It also brings 2x faster reasoning blocks (60 FPS vs 30 FPS), more reliable training, and improved Linux installs and updates. Qwen Image 2.1 G

📉 Far smaller slim image. A slim build now comes down at around 175 MB, near enough 89% smaller than the last release, the local models, the packages around them and the tools that installed them all gone from it; what that changes about the way an instance behaves is set out un

v0.30.0 Highlights This release features 762 commits from 315 contributors (104 new)! New models : DeepSeek-V4.1-Flash ( #56214 , #56228 , #56208 ) with the whole KV stored in MXFP8 through the FlashMLA V4.1 record on SM100 ( #56893 ), DeepGEMM Mega-mHC ( #56962 ), and async…

713 PRs from 237 contributors.

We're releasing our new updated Docker image along with multi-user accounts, RDNA1+2, FP8/INT8 diffusion support, ARM64 CUDA Windows support and many training, GRPO and inference improvements. Also a Qwen3.8-Flash-Next 2x faster MTP hotfix.

Release vllm-proto 0.3.0

Validated by PR #56538 CI at fa2a26f .

Added first-run setup when running ollama, with options to sign in or continue locally. Setup completion is shared with the desktop app on macOS and Windows.

All LiteLLM Docker images are signed with cosign. Every release is signed with the same key introduced in commit 0112e53.

MLX safetensors ollama create no longer experimental. GGUF model creation now requires using llama.cpp tooling for safetensor conversion and quantization.

llama.cpp 0.4.1 adds Maple 20B-A1B, Tencent Hy 4, and Spark2.5 support, improves JSON schema handling, chat parsing, logging, and server child-process management, and updates ggml to v0.24.0.

vllm-proto 0.1.0

All LiteLLM Docker images are signed with cosign. Every release is signed with the same key introduced in commit 0112e53.

72 commits since v0.18.0 (July 17, 2026).

This is a large performance and reliability + bug fix release for Unsloth

This release features 594 commits from 277 contributors (91 new)!

All LiteLLM Docker images are signed with cosign. Every release is signed with the same key introduced in commit 0112e53.

Ollama models can now be used directly in ChatGPT Desktop, so you can keep your existing workflow while running open models. Setup is available from the Ollama app on MacOS.

786 PRs from 214 contributors.

Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2x faster with MTP. MTP is enabled by default, you can still disable it.

This release ships container images only. There is no PyPI package for `1.99.1`.

gemma4 now supports images and audio on MLX engine

A fix-focused release. The main fixes are for few-shot leakage, a multiple-choice filter bug, and group stderr, alongside two new ONNX backends and eight new benchmark suites. Also updated most configs for datasets>=4, which accounts for much of the diff by volume.

♿ Accessibility mode reaches the menus. Accessibility mode now marks the menu entry you are pointing at and the model already chosen with a stronger background, across the dropdown menus, their submenus, and the model picker together with its filter and compare controls, so those

🖼️ Richer previews for terminal files. Word documents and slide decks produced in the terminal are now previewed as the finished document rather than an approximation, and every document preview gains a page strip down the side with numbered thumbnails you can click to jump stra

Ollama's app now follows the system appearance again, restoring dark mode support

Qwen3.8-Flash-Next and GLM-5.3-Flash can now run locally in Unsloth!

MLX: Qwen3.8 Flash Next support

This release features 584 commits from 270 contributors (76 new)!

🚦 Human in the loop tool approval. Where an administrator has turned it on, you can switch a conversation from letting tools run freely to being asked first, so a model that wants to use a tool stops and waits for you to allow or deny it, one call at a time in a saved conversati

Thanks for the support for Qwen3.8-27B and Unsloth Desktop! This is a bug fix release with 170+ PRs.

710 PRs from 212 contributors.

Developers can now easily configure Claude Desktop to seamlessly work with Ollama as a third-party gateway provider.

Thanks for the support for Qwen3.8-27B and Unsloth Desktop last week! For this release, we merged 200+ PRs to introduce many new features, fixes including:

New desktop onboarding flow on first launch

llm: transcode WebP images for llama-server

qwen3.8: support developer instructions

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.

Qwen3.8-27B and Qwen3.8-2.4T can now be run locally in Unsloth!

ollama launch dsh now supports DeepSeek Harness, DeepSeek's open-source agent harness

Unsloth Desktop is here! The first desktop app to run and train AI models locally. Research, export and deploy from the same open-source app on Windows, macOS and Linux.

This is a patch release on top of v0.27.0.

This release features 561 commits from 242 contributors (64 new)!

582 PRs from 194 contributors.

🎨 Redesigned interface. Open WebUI has been visually rebuilt from the ground up. All aspects of the User Interface, from the chat view to the admin panel. Now with a narrower conversation column, lighter typography, tidier spacing, consistent menus and dropdowns, clearly outline

This release features 411 commits from 212 contributors (61 new)!

574 PRs from 169 contributors.

We've been hard at work doing low level improvements in the kernels. Over 90 commits since v0.17.0 (June 3, 2026), themed around fine-tuning very large sparse-MoE models cheaply: 4-bit expert LoRA/QLoRA (NVFP4, MXFP4, bnb) that runs fast and stays memory-flat at long context on B

This release features 2 commits from 2 contributors (1 new)!

v0.5.15.post1 includes a few patches, mostly for GLM 5.2

This release features 558 commits from 232 contributors (64 new)!

GLM-5.2 NVFP4, tuned for production: We took time this cycle to tune GLM-5.2 NVFP4 on Blackwell for optimized production serving. It now runs at 500+ tok/s/user on 8x B300, 450 on 4x GB300 (bs=1). Run GLM-5.2 with our cookbook.

💭 Streamed reasoning display. Models that emit thinking or reasoning now show that content as it streams, and it renders correctly in the chat overview and in exported conversations. Commit, Commit, Commit, Commit

This release features 571 commits from 256 contributors (77 new)!

🤝 Shared folder read-only chats no longer sign users out. Opening or reading chats from shared folders now keeps the current session active when a resource-level access error is returned, instead of incorrectly showing "Session expired. Please sign in again."

🤝 Share folders with your team. You can now share a folder and the chats inside it with specific users, groups, or everyone, with read or write access; people you share with see shared folders in their sidebar and open the chats in a read-only view when they are not the owner, a

New Model Support: GLM-5.2, LiquidAI LFM2.5, Kimi-K2.7-Code, Poolside Laguna-M.1, DiffusionGemma, Zyphra ZAYA1, MiMo-V2-ASR

Stable release of the Continue VS Code extension (final release): removes the CLI-install banner and Generate Rule feature, switches onboarding and the new-config template to explicit model definitions (no Hub slugs), and updates the deprecation banner export link. Re-cut of v1.2

Stable release of the Continue VS Code extension (final release): removes the CLI-install banner and Generate Rule feature, switches onboarding and the new-config template to explicit model definitions (no Hub slugs), and updates the deprecation banner export link.

Please note that Minimax M3 is not yet supported in this version. Please follow vLLM recipe for usage guides for M3.

New Model Support:

This release features 8 commits from 6 contributors (1 new)!

Another packed release. ~84 commits since v0.16.1 last month, bringing Expert Parallelism for MoE training, BitNet 1.58-bit fine-tuning, remote training via Tinker, context parallelism for hybrid SSM models, MXFP4 ScatterMoE-LoRA, fused RMSNorm+RoPE kernels for the Qwen3 family,

📦 Official knowledge base sync tool. A new companion tool from Open WebUI, oikb, keeps a knowledge base in sync with a local directory, GitHub repo, S3 bucket, Confluence space, or any of more than 40 other sources, uploading only new and changed files using the incremental sync

v0.5.12.post1 is a stability patch on top of v0.5.12. It cherry-picks 12 fixes — primarily for DeepSeek V4 — onto the release branch.

DeepSeek V4 support: Full inference path for DeepSeek-V4 (#23882), including:

New release with four new model backends, tensor parallel support for transformers based models (hf), new benchmarks, a TaskManager refactor, and a long tail of task correctness fixes.

🛡️ Redirect-based SSRF protection. All outbound HTTP requests now block 3xx redirects by default via a new AIOHTTP_CLIENT_ALLOW_REDIRECTS environment variable, preventing redirect-based SSRF where a public URL silently redirects to internal addresses (RFC 1918, loopback, cloud-m

📜 Chat scroll position on load. Opening a chat conversation now reliably scrolls to the bottom of the message history, fixing a regression caused by content-visibility: auto where estimated element sizes prevented the initial scroll from reaching the true bottom.

Lots of bugfixes

Caching system prompt and user messages for non-trimmable caches

Example YAML: https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/gemma4/26b-a4b-moe-qlora.yaml

We’re very excited to share this new packed release. We had ~80 new commits since v0.15.0 (March 6, 2026).

fix: ensure config.yaml exists and is populated when accessed by @RomneyDa in https://github.com/continuedev/continue/pull/11915

Fix save/load of CacheList by @angeloskath in https://github.com/ml-explore/mlx-lm/pull/886

This release brings new model support, significant MoE improvements, infrastructure updates with Torch 2.10.0 and uv builds, and a collection of quality-of-life fixes across the board.

Minor release. Stay tuned for bigger changes next release.

Fix Kimi Linear by @kernelpool in https://github.com/ml-explore/mlx-lm/pull/853

Transformers v5 by @awni in https://github.com/ml-explore/mlx-lm/pull/811

This is a major release marking our migration to Transformers v5. Along with this significant core dependency upgrade, we are introducing major performance optimizations for MoE models and new fine-tuning methods.

The big change this release: the base package no longer installs model backends by default. We've also added new benchmarks and expanded multilingual support.

import logging as it throws no logging error in place of actual error by @Maanas-Verma in https://github.com/ml-explore/mlx-lm/pull/778

This is a patch release introducing GDPO support and updating core infrastructure, including newer CUDA defaults and Python versions.

This release brings support for PyTorch 2.9.1, expands our ecosystem with new experiment trackers (SwanLab and Trackio), and introduces support for a wide range of new models including Olmo3, Ministral 3, InternVL 3.5, and Kimi. We’ve also included significant improvements to qua

Add AWQ/GPTQ weight transformation utilities by @ericcurtin in https://github.com/ml-explore/mlx-lm/pull/730

Fix mlx-lm release by @awni in https://github.com/ml-explore/mlx-lm/pull/733

custom dsv32 chat template by @awni in https://github.com/ml-explore/mlx-lm/pull/693

fix: server busy-waiting during idle request polling by @zenyr in https://github.com/ml-explore/mlx-lm/pull/674

This release is packed with major new features, including Streaming SFT for massive datasets, a new Text Diffusion training plugin, and a significant upgrade to our Quantization-Aware Training (QAT) capabilities with NVFP4 support. We're also thrilled to announce support for a hu

This release continues our steady stream of community contributions with a batch of new benchmarks, expanded model support, and important fixes. A notable change: Python 3.10 is now the minimum required version.

Added support for all GPT-5 models.

This v0.4.9.1 release is a quick patch to bring in some new tasks and fixes. Looking aheas, we're gearing up for some bigger updates to tackle common community pain points. We'll do our best to keep things from breaking, but we anticipate a few changes might not be fully backward

Added support for new Gemini models including gemini-2.5-pro, gemini-2.5-flash, and gemini-2.5-pro-preview-06-05 with thinking tokens support.

Enhanced Backend Support:

Added support for new Claude models including the Sonnet 4 and Opus 4 series (e.g., claude-sonnet-4-20250514,

Automated build release.

New multimodal engine.

Add Qwen3 and Qwen3MoE support

Added support for gemini-2.5-pro-preview-05-06 models.

Add Torch 2.7.0 wheels (big thanks to @kingbri1 for unborking the build action)

Support for GPT 4.1, mini and nano.

Added support for the openrouter/openrouter/quasar-alpha model.

OpenRouter OAuth integration:

Added support for SOTA Gemini 2.5 Pro.

Added support for thinking tokens for OpenRouter Sonnet 3.7.

Big upgrade in programming languages supported by adopting tree-sitter-language-pack.

New Backend Support:

Support Qwen2.5-VL

Re-architected core engine with lower CPU overhead.

Basic video support for Qwen2-VL

This release includes several bug fixes, minor improvements to model handling, and task additions.

Some small fixes, most notably for Qwen2-VL inference on Windows

Initial support for Qwen2-VL (images for now, no video)

This release brings important changes to chat template handling, expands our task library with new multilingual and multimodal benchmarks, and includes various bug fixes.

Support Pixtral

No longer use safetensors for loading weights (fix virtual memory issues on Windows especially)