Change8

v0.1.701-beta

📦 unslothView on GitHub →
13 features🐛 4 fixes

Summary

Unsloth Desktop is released, a new app for running and training AI models locally on various platforms and hardware. It introduces features like enhanced tool calling, support for new models, faster inference, no-code training, and remote deployment.

✨ New Features

  • Introduction of Unsloth Desktop app for running and training AI models locally on Windows, macOS, and Linux.
  • Up to 50% more accurate tool calling with self-healing calls and sandboxed code execution.
  • Support for running models like Muse Glimmer 30B, Kimi K3, Qwen3.8, DeepSeek-V4 Flash 0731, and Gemma 4.
  • Video generation with MiniMax-H3 and up to 2x faster inference for image and video diffusion models.
  • Unlimited private web search, Deep Research for cited reports, and RAG capabilities.
  • Export models to NVFP4, GGUF, and other formats.
  • Remote access to Unsloth models via Cloudflare HTTPS.
  • Support for running on CPU or multiple GPUs across NVIDIA, AMD, Intel, and Mac hardware.
  • No-code model training with reduced time and VRAM usage.
  • Local models served through an OpenAI-compatible API.
  • Integration with OpenAI and Anthropic models as cloud providers.
  • Sandboxed Python and Bash execution for models to test code and create files.
  • Support for LoRAs, reference images, and ControlNet in diffusion model workflows.

🐛 Bug Fixes

  • Fixed slow Windows downloading, now 200x faster due to throttling.
  • Fixed Mac issue where it prompted to download command line tools (resolved uv bug).
  • Fixed AMD Strix Halo not being detected.
  • Other unspecified bug fixes.