●Change8

v0.1.804-beta

📦 unslothView on GitHub →
✨ 35 features🐛 38 fixes🔧 55 symbols

Summary

Unsloth now supports running Qwen3.8-Flash-Next and GLM-5.3-Flash locally with significant performance improvements and reduced memory requirements. This release also brings numerous chat, reliability, and model management enhancements.

Migration Steps

  1. Bump install.sh / install.ps1 pin to unsloth>=2026.8.21

✨ New Features

  • Qwen3.8-Flash-Next and GLM-5.3-Flash can now run locally.
  • Run Qwen3.8-Flash-Next on 75GB RAM.
  • GLM-5.3-Flash runs on 102GB of combined RAM + VRAM.
  • 5x faster inference than `--fit on` if RAM offloaded.
  • 100+ chat, reliability and performance improvements.
  • Qwen3.8-Flash-Next is a new 125B multimodal reasoning model.
  • The 1-bit Unsloth Dynamic GGUF for Qwen3.8-Flash-Next runs on 75GB RAM or unified memory.
  • Qwen3.8-Flash-Next supports chat with text and images using up to 262K context.
  • Qwen3.8-Flash-Next allows switching between None, Low, Medium and Extra High reasoning.
  • Preserved Thinking in Qwen3.8-Flash-Next keeps reasoning consistent across longer chats.
  • GLM-5.3-Flash is Z.ai's new 320B multimodal model, with only 18B parameters active at a time.
  • The 1-bit GLM-5.3-Flash model runs on 102GB of combined RAM + VRAM.
  • GLM-5.3-Flash supports chat with text, images and long documents using up to 1M context.
  • GLM-5.3-Flash allows switching between Low, High and Max reasoning.
  • GLM-5.3-Flash offers stronger coding, agent and vision performance than GLM-5.2.
  • Recommended settings are applied automatically for GLM-5.3-Flash in Unsloth.
  • Local chats resume after a disconnect instead of losing the reply.
  • Deep Research keeps going when a provider asks it to slow down.
  • Vision chats now handle multiple images properly.
  • Images returned by MCP tools appear directly in chat.
  • Export chats as JSONL for backups or use in other tools.
  • Adjust Auto Compaction for longer chats, or turn it off.
  • Collapse tool activity by default for cleaner agent chats.
  • Large GGUFs automatically split across GPU and system RAM.
  • See estimated memory usage before loading a model.
  • View VRAM usage directly from your downloaded models.
  • Model settings stay saved when switching chats.
  • Search and download embedding models directly from Hugging Face.
  • Text-to-speech models only load when you actually use them.
  • Linux voice recording fixed.
  • NVIDIA + Wayland interface freezes fixed.
  • AMD model loading crashes fixed.
  • llama.cpp models now load from Windows profiles with non-English characters.
  • Non-English web links now work properly as chat sources.
  • Desktop download links always point to the latest stable release.

🐛 Bug Fixes

  • Linux voice recording fixed by capturing raw PCM where WebKitGTK's MediaRecorder produces no audio.
  • Studio: resolve PowerShell by absolute path in the update gate.
  • Prevent shared preview loads from evicting the active Studio model.
  • Studio: harden lockfile audit followups for #5604.
  • Studio: take remend 1.3.1, which stops repairing markdown that is already complete.
  • CLI: follow model reasoning default in unsloth start.
  • Name the encoding when the poll probe writes its artifact.
  • Studio: add a freeze report script for Linux desktop interface hangs.
  • Fail the mac tab-capability smoke on a backend that never comes back, not on a stall.
  • Warm the app before timing /api/health, instead of widening the ceiling.
  • docs: remove duplicate Search bullet and fix error messages.
  • Follow up #9435: restore the rest of the moved names, and read quant labels through the variant lister.
  • Studio: use --load-mode none when the model fits without paging.
  • Studio: spill weights with -ot instead of layers, so the KV cache stays on the GPU.
  • Reject a negative chunking stride instead of silently dropping text.
  • Studio: fix web pages with non-English URLs failing to load.
  • Studio: align composer and research activity icons.
  • desktop: turn webkit compositing off on the one wayland nvidia setup that freezes.
  • Studio: keep a resumed turn's provider metadata through the continuation merge.
  • Studio: render MCP images returned as embedded resources.
  • Studio: send the most recent chat image to vision models.
  • Studio: align sidebar list header actions, and mark downloaded models with a dot.
  • Prevent stale hydration from overwriting remembered model settings.
  • Handle invalid Krea2 model indexes clearly.
  • Studio: track API token usage in Profile.
  • Studio: tell the model which documents are attached.
  • Studio: use live trending GGUFs in agent setup.
  • studio: preserve gguf settings across chat switches.
  • studio: add message jsonl export.
  • Validate studio_install_id before baking it into the launcher.
  • studio: show lm studio model names in chat.
  • Studio: fix per-chat settings and fork counts for app-created chats.
  • studio: show the train run preview as a column on a laptop window.
  • Use Unsloth for bare Studio product references.
  • Studio: stop the temporary toggle tagging a saved app-created chat.
  • Studio: honour forced tool_choice on local GGUF tool loops.
  • Desktop: publish stable latest downloads.
  • Give every installing job its own pip cache key.

Affected Symbols

install.shinstall.ps1unslothPowerShellWebKitGTKMediaRecorderremendunsloth startpoll probemac tab-capability smoke/api/healthSearch bulletquant labelsvariant lister--load-mode none-otKV cachechunking strideMCP toolscomposer activity iconsresearch activity iconswebkit compositingWaylandNVIDIAprovider metadataembedded resourcesvision modelssidebar list header actionsdownloaded modelshydrationKrea2 model indexesAPI tokenProfiledocumentstrending GGUFsagent setupgguf settingschat switchesmessage jsonl exportstudio_install_idlauncherlm studio model namesper-chat settingsfork countsapp-created chatstrain run previewlaptop windowStudio product referencestemporary togglesaved app-created chatforced tool_choicelocal GGUF tool loopsstable latest downloadsinstalling jobpip cache key