unsloth Releases

This is a large performance and reliability release featuring significant speed improvements for diffusion and inference, especially on AMD and Apple Silicon, alongside over 250 bug fixes and a 60% reduction in binary size.

Sep 9, 2026

This release focuses on performance and reliability improvements for AMD and Windows users, alongside a major dependency bump to PyTorch 2.11. It introduces new Docker images, fixes over 200 bugs, and enhances Studio features including DoRA training on Apple Silicon and improved attachment handling.

Sep 8, 2026

Significantly boosts inference speed for Qwen and GLM models using MTP, introduces new audio and video generation APIs, and enhances MLX support.

Sep 2, 2026

This release introduces Model Training Protocol (MTP) to enable up to 2x faster inference for Qwen3.8-Flash and GLM-5.3-Flash models. It features 170+ updates across training, chat, and hardware, including new audio models, improved AMD/ROCm support, and significantly faster MLX inference on Apple Silicon.

Sep 2, 2026

Added support for Qwen3.8-Flash-Next and GLM-5.3-Flash models, enabling local inference with specific RAM requirements. Includes infinite repeated compaction, chat recovery after disconnects, and fixes for AMD model loading and Linux voice recording.

Aug 27, 2026

A bug fix release featuring MLX/Mac fixes, LAN API keyless access, and AMD bug fixes. It introduces Auto Compaction (Experimental) and LAN Remote Access (Preview).

Aug 25, 2026

Identical to v0.1.803-beta. A bug fix release featuring MLX/Mac fixes, LAN API keyless access, and AMD bug fixes. It introduces Auto Compaction (Experimental) and LAN Remote Access (Preview).

Aug 25, 2026

Introduced Auto Compaction (Experimental) and LAN Remote Access (Preview). Features faster chat, custom llama.cpp builds, and Unsloth Dynamic v3.0 support.

Aug 20, 2026

Added support for Qwen3.8-27B and Qwen3.8-2.4T. Introduced MiniMax-H3 video generation, custom llama-server arguments, and improved AMD/Intel/Mac support.

Aug 14, 2026

Launch of Unsloth Desktop. Added tool calling, web search, and Codex login. Fixed AMD Strix Halo detection and Windows download throttling.

Aug 13, 2026

Identical to v0.1.70-beta. Launch of Unsloth Desktop. Added tool calling, web search, and Codex login. Fixed AMD Strix Halo detection and Windows download throttling.

Aug 11, 2026

Identical to v0.1.701-beta. Launch of Unsloth Desktop. Added tool calling, web search, and Codex login. Fixed AMD Strix Halo detection and Windows download throttling.

Aug 11, 2026

Aug 11, 2026

Introduced Meta Muse Glimmer 30B. Added MiniMax-H3 video generation support, preliminary image diffusion, revamped training page, and downloadable chat artifacts.

Aug 10, 2026

Identical to v0.1.61-beta. Introduced Meta Muse Glimmer 30B. Added MiniMax-H3 video generation support, preliminary image diffusion, revamped training page, and downloadable chat artifacts.

Aug 10, 2026

A maintenance release focusing on improvements to the Unsloth Studio and Desktop applications, including independent prompt queues for parallel chats, sandboxed matplotlib rendering, and various UI/UX fixes.

Aug 9, 2026

Introduces Docker image signature verification for LiteLLM images using cosign, allowing users to verify the integrity of the v1.94.2 Docker image.

Aug 8, 2026

Identical to v0.1.525-beta, focusing on bug fixes, Mac improvements, and the default DSpark support for faster inference.

Aug 4, 2026

Focuses on bug fixes and Mac-specific improvements alongside the existing DSpark and model support, aiming for smoother installations.

Aug 4, 2026

Enhances previous releases by enabling DSpark support by default for DeepSeek V4 Flash, offering 2x faster inference, while maintaining support for Kimi K3 and optimized downloading.

Aug 4, 2026