●Change8

v0.1.806-beta

📦 unslothView on GitHub →
✨ 40 features🐛 27 fixes⚡ 1 deprecations🔧 26 symbols

Summary

This release significantly enhances performance for Qwen and GLM models with MTP, introduces new local media APIs, and improves audio capabilities. It also brings numerous bug fixes and performance optimizations across various features like MLX inference and chat functionalities.

Migration Steps

  1. Put the smart offload planner back behind its flag if needed.

✨ New Features

  • Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2x faster with MTP (enabled by default).
  • Smoother model loading with fewer errors across local servers and connected providers.
  • Safer chat edits that preserve tool cards, reply details, and conversation branches.
  • New local media APIs for video, audio, and MLX-served models.
  • New audio support with models like MiniMax-Music3, Higgs, MOSS, and progress tracking.
  • Improved multi-GPU planning, memory fitting, and split-model training.
  • Ctrl+F Search is now functional.
  • Strengthened AMD/ROCm detection, installation, and GPU compatibility.
  • Upgraded MCP, Deep Research, and OAuth reliability.
  • Use GLM tools across longer, multi-turn chats.
  • Qwen automatically applies recommended settings for thinking and non-thinking modes.
  • Fine-tune large MoE models with text or images on Apple Silicon using MLX.
  • Long Qwen chats run much faster on Mac with follow-up turns up to 30x faster.
  • MLX models now use their full context size and support much longer batched generation.
  • MLX models release GPU memory more cleanly between generation bursts and model switches.
  • Serve MLX models through Unsloth's OpenAI-compatible API.
  • Audio clips can now be archived and managed.
  • Run multiple tool calls simultaneously without argument mix-ups.
  • Keep tools available when chatting with images.
  • Each chat maintains its MCP connection for faster tool calls.
  • Local models can edit code using Codex's apply_patch tool.
  • Continue long chats with images and other media using Auto Compaction.
  • Review and approve Deep Research plans before research begins.
  • Train larger models across multiple GPUs with automatic placement.
  • AMD installs select the best build across Windows and Linux, with BF16 on more GPUs.
  • Export GLM-5.3 MLX fine-tunes to GGUF.
  • Choose custom GGUF shard sizes and save locations.
  • Generate videos through the new OpenAI-compatible Videos API.
  • Updates download in the background and install upon restart.
  • Choose a custom port for LAN access.
  • Model downloads show clearer progress and can switch from Xet to HTTP automatically.
  • Added support for MiniMax-Music3, Higgs, MOSS audio models.
  • Added live progress updates while audio is being generated.
  • Improved reliability with custom TTS playback fixes, Whisper pairing checks, and stronger audio testing.
  • Generate audio with Higgs, MOSS and MiniMax models.
  • Track audio generation progress and archive finished clips.
  • Studio: serve installed mlx models through the openai-compatible api.
  • Studio: add the OpenAI videos API on /v1/videos.
  • Pin Claude Code to the loaded model's real context window.
  • Studio: list route-aware models through the openai api.

🐛 Bug Fixes

  • Desktop build completion issue due to mismatched linuxdeploy-plugin-appimage digest fixed.
  • Fix Studio hydrating synced GGUF files before selection.
  • Fix datasets PyArrow registration after failed warm import.
  • Fix Studio CPT overwriting LFM2 all-linear LoRA targets.
  • Fix Studio accepting trailing slash for model discovery.
  • Studio: tell users when a canvas is blocked by the network access setting.
  • Studio: fix streamed tool calls in studio api monitor.
  • Fix sidebar footer width and context menu radius.
  • Strip inherited claude provider routing from local agent launches.
  • Studio: honor a cancelled load inside the llama-server health wait.
  • Fix model hub recent sorting for on-device models.
  • Studio: fix lan address actions and launch bind reporting.
  • RAG: avoid re-hashing linked-folder snapshots in start_ingestion.
  • Studio: coalesce concurrent keyless-settings reads to stop SQLite thundering herd.
  • Clarify tool approval copy.
  • Fix shell selection for remote Agents commands.
  • Studio: accept the loadConfig keys a saved system-prompt preset already sends.
  • Studio: compact long chats with media attachments.
  • Studio: add Higgs, MOSS, and MiniMax audio models.
  • Studio: add ordered arguments to mcp server settings.
  • Studio: repair custom TTS endpoint test and playback.
  • Studio: render escaped inline math in lists.
  • Restore UNSLOTH_RETURN_LOGITS after prediction_step instead of forcing it to 0.
  • Studio: stop the per-chunk autosave writing back messages the server owns.
  • Fix Studio accepting shared youtube clipboard links.
  • Fix Studio not offering a transformers upgrade where it cannot load anything.
  • Studio: drop the schema bounds llama.cpp's grammar engine cannot compile.

Affected Symbols

linuxdeploy-plugin-appimageQwen3.8-Flash-NextGLM-5.3-FlashMLXMiniMax-Music3HiggsMOSSWhisperCodexClaude Codellama.cppPyArrowtransformersLFM2GGUFXetHTTPMCPOAuthROCmBF16OpenAI-compatible APIVideos APIinstall.shinstall.ps1unsloth>=2026.8.22

⚡ Deprecations

  • Unsupported custom prompt templates are deprecated.