v0.1.800-beta

✨ New Features

• Qwen3.8-27B and Qwen3.8-2.4T can now be run locally in Unsloth, including fine-tuning Qwen3.8-27B.
• Run on 17GB RAM via Unsloth Dynamic GGUFs.
• NVFP4 quants uploaded for Qwen3.8-27B.
• Extra llama-server arguments allowed with a custom VRAM toggle.
• External provider now supports tool calling, tool support, and login with Codex.
• Fast FP8 10x faster MiniMax-H3 inference.
• 10% faster inference for GGUFs.
• Bypass permissions fixed for GGUFs.
• Connected AI providers can use their own Search or Unsloth Desktop's built-in Search and tools, with tool results passed back to the model for multi-step tasks.
• Sign in with a Codex subscription to use Codex tools inside Chat.
• Chat shows live prompt and generation speeds, and long streaming replies use less CPU.
• Chat settings persist across remote sessions.
• Paste a YouTube link to attach its transcript (title, channel, duration, link, caption language).
• Save full chats or replies into project sources, preserving reasoning, tool calls, and citations.
• MiniMax-H3 can run on smaller supported GPUs by splitting large model parts.
• Model picker hides H3 options that current hardware cannot run.
• H3 options are labeled 'Fast FP8' or 'Slow' to indicate speed differences.
• Inference is up to 10% faster with lower VRAM use and a tunable memory limit.
• Idle image and video models can optionally unload to free VRAM for Chat or Training.
• Added better support for AMD RDNA 3, RDNA 4, and Strix Halo systems.
• VRAM checks no longer reserve extra GPU memory.
• Multi-GPU ROCm device matching is safer.
• Macs now choose context size from actually free memory.
• RAG document indexing uses the CPU by default, freeing up GPU memory.
• Model settings include an 'Extra Arguments' box for custom `llama-server` flags.
• Unsloth checks custom `llama-server` flags against the installed build and saves valid ones per model.
• Training checks for newer Transformers versions and prompts before installing.
• Training preview indicates when a model must use 16-bit instead of 4-bit.
• GGUF exports check disk space before starting a merge and provide clearer error messages.
• On Kaggle, large saves can use the bigger temporary drive.
• Settings > Debugging now shows main server and model-runner logs with live refresh modes.
• Copy visible log or full path for bug reports, with sensitive information hidden in the viewer.
• Cached RAG models load first to avoid stalls when Hugging Face is unreachable.
• Slow installs can continue during large downloads and show progress.
• Pinned models can be reordered by dragging in the model list.