●Change8

v0.1.801-beta

📦 unslothView on GitHub →
✨ 18 features🐛 20 fixes🔧 39 symbols

Summary

This release introduces experimental Auto Compaction for longer chats and a preview of Remote & LAN Access. It also brings significant speed improvements to chat, support for custom llama.cpp builds, and enhanced features for the Unsloth Dynamic GGUFs.

Migration Steps

  1. Change generated admin password for Remote & LAN access.
  2. Update install.sh / install.ps1 to pin to unsloth>=2026.8.18.

✨ New Features

  • Introduced Auto Compaction (Experimental) for longer chats beyond context limits, allowing evicted turns to remain searchable.
  • Added Remote & LAN Access (Preview) for easy network access without Cloudflare links.
  • Improved streaming performance, reduced UI lag, and smoother long conversations for faster chat experiences.
  • Added support for custom llama.cpp builds with toggles for Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, and Vision On/Off.
  • Released Unsloth Dynamic v3.0 GGUFs for Qwen3.8-27B, delivering >10% higher top-1 accuracy.
  • Added a dedicated Settings section for remote access, supporting connection addresses, QR codes, and optional auto-start.
  • Projects feature to organize chats, files, and workspaces together.
  • Chats now remember composer settings, system prompts, and model sampling options.
  • Added prompt queueing and customizable keyboard shortcuts.
  • Added `edit_file` tool for partial file modifications.
  • Improved tool calling and MCP behavior.
  • Added Intel XPU support.
  • Responses API supports structured output formatting.
  • Added OpenCode V2 support for `unsloth start`.
  • Hub supports more image/video workflows.
  • Faster and improved GGUF downloads.
  • Desktop packaging and permissions improved.
  • Training scheduler and preprocessing improvements.

🐛 Bug Fixes

  • Fixed dead Linux .deb download link.
  • Studio: preserved pre-tool reasoning inside the GGUF tool loop.
  • Studio: sped up the local model inventory and kept it off the API loop.
  • Studio: reused cached GGUF work for load requests.
  • Studio: kept each tool-call round separate when the delta index restarts.
  • Studio: routed bare tool-call fragments to the call that owns the index.
  • Studio: fixed preprocessing of rows for max_steps runs to only use what's needed.
  • Studio: stopped the chat UI from falling behind fast streams.
  • Studio: kept each chat's composer pills and settings with the chat.
  • Crash guard: read dumpability in order, and stopped losing crashes to scope and naming.
  • Studio: sized the max_steps row bound for the data-parallel world.
  • Studio: made the paste-to-attachment threshold a chat setting.
  • Fixed CI on main: stale test doubles, a stale router stub, and two source defects.
  • Stopped six backend tests waiting on the clock instead of on a signal.
  • CI: stopped three workflows paying for work they throw away.
  • Kept xFormers working when flash-attn 4 is installed, and guarded the varlen int32 overflow.
  • Improved recovery when llama-server crashes.
  • Better validation for missing GPU support and oversized GGUF models.
  • Improved MTP performance and VRAM handling.
  • Better compatibility across ROCm, xFormers, and flash-attention setups.

Affected Symbols

llama.cppunsloth.ai/docs/basics/dynamic-3.0-ggufssearch_conversationedit_fileunsloth startinstall.shinstall.ps1llama-flag fixturesparallelSlotsClampedGGUF tool looplocal model inventoryAPI loopcached GGUF worktool-call rounddelta indexbare tool-call fragmentsimage and video apisProjects sidebarRecentsDocuments folderproject workspacesmax_steps runvalidated datasetschat UIcomposer pillssystem promptsmodel sampling optionsper-connection max tokens limitCrash guarddumpabilityscope and namingstartup profilexFormersflash-attn 4varlen int32 overflowbackend testsworkflowsROCmflash-attention