Unsloth Desktop - Open-source app unifying local AI model execution, training, and agents

Unsloth has released Unsloth Desktop beta, an open-source app available on macOS, Windows, Linux, and WSL. It allows users to download models and chat locally without separate configuration, handling execution, training, and deployment in one place.

Supported targets include LLMs, diffusion image/video models, MLX, GGUF, and audio models. Users can select and download quantized models suited to their device performance from the model hub, and existing model folders are automatically recognized.

Key features include:

• Agent integration: Connects with Claude Code, Codex, MCP, web search, etc., and links local models to agents via the unsloth start command
• Tool execution security: Controls file and internet access permissions, running Bash and Python in a sandbox
• Model training: Supports LoRA, full fine-tuning, and pre-training for text, diffusion, embedding, and audio models
• Media generation: Supports image/video generation and editing, as well as LoRA training for FLUX, Z-Image, LTX, Wan, etc.
• Voice processing: Generates, fine-tunes, and transcribes locally using TTS, STT, Whisper, Qwen3-ASR, etc.
• Deep research: Creates a search plan first, then generates reports with sources and citations
• Model serving: Makes local models accessible externally via Cloudflare tunnels or network binding
• Cloud models: Uses OpenAI, Anthropic, Ollama, llama.cpp, vLLM, and other APIs within the same chat interface

Unsloth explains that it provides a self-healing tool call feature that detects tool call failures, fixes them, and retries, aiming for up to 50% more accurate tool calls. Regarding training capabilities, it stated that it supports up to 2x faster training and 70% less VRAM usage based on existing Unsloth technology.

The app has no telemetry and supports fully offline execution. It supports NVIDIA, Intel, and AMD GPUs, Apple devices, and CPUs, though older hardware may be limited.