Unsloth publishes an open-source desktop application that combines execution, image and video generation, and fine-tuning of models on its own machine, on macOS, Windows and Linux.
Unsloth released on August 11, 2026 a beta version of Unsloth Desktop, a free desktop application that brings together in a single interface the execution and training of models on personal hardware. It runs on macOS, Windows and Linux, including via WSL, and accepts NVIDIA, Intel, AMD and Apple chips.
The software supports several families of formats: GGUF for language models, MLX Apple side, as well as image and video streaming models (SDXL, FLUX.2) and audio models. Recent catalogues can be accessed directly from Qwen3.8 to DeepSeek V4, Kimi K3 or Meta's Muse Glimmer.
Gains announced on inference and tooling
Unsloth claims an improvement of about 50% in the accuracy of tool calls, as well as significantly reduced video generation times on high-end hardware — from just over 70 seconds to 13 seconds for a sequence of 124 images in 960×544 on B200 with MiniMax-H3 in FP8. The application also enables you to connect agents like Claude Code or Codex to a locally executed model, and expose a service via a Cloudflare tunnel. Built with Tauri, it relies on llama.cpp, PyTorch and Hugging Face infrastructure.
Read the full article on the original source
Go to the source →