Unsloth Studio
A local UI for running and fine-tuning LLMs with up to 2x faster training and 70% less VRAM.
Evidence shows: default binds to localhost only, remote access requires explicit options; provides --secure and --disable-tools security options; documentation clearly states code execution and tools are on by default and run as your user, with warnings about exposure; password setting supports env var and stdin to avoid CLI leakage; dependencies have version pins and conditional constraints; external effects (e.g., Cloudflare tunnel) are clearly explained with fail-closed mechanism; provides uninstall scripts and rollback paths; source attribution is clear (authors, maintainers, copyright). Deductions: publisher identity is unverified, and some security measures rely on user opt-in, not default least privilege.
Evidence shows: README and pyproject.toml descriptions are consistent; feature list corresponds to documentation links; dependencies have version ranges and conditional constraints, but some (e.g., torchcodec) have platform restrictions that may cause install failures; documentation provides troubleshooting and error handling (e.g., cache cleanup, CI failure handling). Deductions: no specific runtime error message examples, and dependency availability not tested.
Evidence shows: clearly targets developers and researchers, supports multiple OS and hardware (NVIDIA, AMD, macOS, Vulkan); capability boundaries are clear (training, inference, export); triggers are precise (env vars, CLI options); environment fit is good (Docker, install scripts, platform-specific guides). Deductions: some advanced features (e.g., MCP, code execution) require user configuration, no default security boundary.
Evidence shows: README structure is clear with features, install, usage, news; install notes are detailed with multiple methods; naming is stable (unsloth command); provides examples and FAQ (via doc links); known limitations are partially mentioned (e.g., Vulkan inference only); license is clear (Apache-2.0, but studio part is AGPLv3); versioning via CHANGELOG and dynamic version; maintenance responsibility is clear (authors and team). Deductions: no full CHANGELOG content, and dual licensing may cause confusion.
Evidence shows: output usability is high (multiple export formats and APIs); marginal value is clear (faster training, less VRAM); cost-benefit is reasonable (free open source, free notebooks). Deductions: performance claims lack benchmark data, and actual results depend on hardware and configuration.
Evidence shows: README claims are supported by documentation links; multiple sources (README, pyproject, workflows) corroborate; facts and inferences are separated (e.g., performance claims marked 'up to'). Deductions: some performance numbers lack specific test methodology, and no reproducible benchmarks provided.
- Publisher identity is unverified; treat with caution.
- Code execution and tools are on by default; use --disable-tools and protect API keys when exposing the service.
- Dual licensing (Apache-2.0 and AGPLv3) may affect compliance.
- Some dependencies (e.g., torchcodec) have platform restrictions that may cause install failures.
- Performance claims lack benchmark data; actual results may vary by hardware.
What does this agent do, and when should you use it?
Unsloth is an open-source project offering two ways to run and fine-tune large language models locally: Unsloth Studio, a web UI, and Unsloth Core, a code library. It supports text, audio, embedding, and vision models, and runs on NVIDIA, AMD, Apple Silicon, and Vulkan devices. Unsloth's optimized kernels and packing algorithms claim up to 2x faster training (12x for MoE) with 70% less VRAM. Inference features include model search and download (GGUF, LoRA, safetensors), tool calling, code execution, web search, and OpenAI/Anthropic-compatible API endpoints. Training supports LoRA/QLoRA, full fine-tuning, reinforcement learning (GRPO), and long-context training, plus data recipes that auto-create datasets from PDF/CSV/DOCX. Commands like `unsloth start claude` connect local models to coding agents, and the project is dual-licensed under Apache 2.0 and AGPL-3.0.
Unsloth Studio is a local web UI for browsing, downloading, and running models (GGUF, LoRA, safetensors), with chat, tool calling, code execution, and web search. It exposes OpenAI/Anthropic-compatible APIs (/v1/chat/completions, /v1/responses, /v1/messages) to let external clients use local models, and unsloth start integrates with Claude Code, Codex, and other agents. For training, it fine-tunes 500+ models using custom Triton kernels for speed and memory efficiency, supports reinforcement learning (GRPO) and long contexts, and offers data recipes to create datasets from documents. Installation is via shell command (curl or PowerShell), and usage starts the server on a local port with optional Cloudflare HTTPS tunneling for remote access.
- A data scientist wanting to fine-tune Llama 3.1 locally with 70% less VRAM, using free Colab notebooks.
- A developer needing a local model backend for Claude Code or Codex, connecting with
unsloth start claude. - A researcher experimenting with reinforcement learning (GRPO) on Qwen3, leveraging Unsloth's claimed 80% VRAM reduction.
- A user running GGUF models with hardware controls like MoE expert offloading and multi-GPU tensor parallelism.
- A team automating dataset creation from PDFs and CSVs using Unsloth's data recipe feature with visual node editing.
- A macOS user wishing to run and fine-tune models on Apple Silicon via MLX support.
What are this agent's strengths and limitations?
- Significant speedups (up to 12x for MoE) and memory savings (70% less VRAM) with long-context training (500K).
- Provides both a graphical web UI and a code library, catering to different expertise levels.
- Broad hardware compatibility (NVIDIA, AMD, Apple Silicon, Vulkan) and cross-platform install scripts.
- OpenAI/Anthropic-compatible APIs and
unsloth startsimplify integration with coding agents. - Supports a wide range of models (Llama, Qwen, Gemma, DeepSeek, etc.) with ongoing updates.
- Unsloth Studio UI is AGPL-3.0 licensed, which may deter organizations requiring permissive licensing.
- Training depends on PyTorch and requires a compatible GPU; CPU-only users are limited to chat and data recipes.
- Installation and configuration can be complex, especially for custom setups or Vulkan backend.
- Some advanced features require substantial VRAM (e.g., 80GB for 500K context), limiting consumer hardware.
- Beta status means potential instability or frequent changes.
How do you install or deploy this agent?
For macOS/Linux/WSL, run curl -fsSL https://unsloth.ai/install.sh | sh. For Windows PowerShell, run irm https://unsloth.ai/install.ps1 | iex. To force the Vulkan backend, set UNSLOTH_FORCE_VULKAN=1 before installing. For Unsloth Core, create a Python 3.13 venv with uv and run uv pip install unsloth --torch-backend=auto. Docker users can use unsloth/unsloth image.
How do you use this agent?
After installation, launch the web UI with unsloth studio -p 8888. It binds to localhost by default; use -H 0.0.0.0 for LAN access or --secure for a Cloudflare HTTPS tunnel. In the UI, search and download a model, then choose chat or training mode. For training, upload a dataset or use data recipes to create one, and configure parameters like LoRA/QLoRA or RL. To connect a coding agent, use unsloth start claude, unsloth start codex, etc., or as a subagent with unsloth start claude --as-subagent --model unsloth/model-GGUF:quant.
FAQ
Is Unsloth free?
How does Unsloth compare to llama.cpp?
Can I do reinforcement learning with Unsloth?
How do I connect a local model to Claude Code?
unsloth start claude after launching Unsloth and loading a model. It supports using Claude Code as a subagent.