Visão geral
Unsloth provides two main interfaces: Unsloth Studio, a web UI for local model workflows, and Unsloth Core, a code-based package. Studio supports model discovery, inference, dataset preparation, training, reinforcement learning, observability, export, and compatible API endpoints. The repository also describes integrations for local agents, MCP clients, external API providers, vLLM, and Ollama.
Principais recursos
- Beta web UI for local inference, training, dataset preparation, monitoring, and export
- Code-based Unsloth Core package
- Support for text, audio, embedding, and vision workflows
- GGUF, LoRA adapter, and safetensors model handling
- LoRA, QLoRA, full fine-tuning, reinforcement learning, pretraining, 4-bit, 16-bit, and FP8 workflows
- OpenAI-compatible /v1/chat/completions and /v1/responses endpoints
- Anthropic-compatible /v1/messages endpoint
- Agent integration through the unsloth start command
- Opt-in MCP endpoint for controlling models, training, recipes, checkpoints, and exports
- Web and PDF search, tool calling, and sandbox or artifact code execution
- Model comparison using the same prompt
- Multi-GPU and GGUF hardware-placement controls
- Training observability for loss and GPU usage
- Connections to OpenAI, Anthropic, vLLM, and Ollama
Requisitos, instalação e início rápido
Uso
Compatibilidade do modelo e casos de uso
The supplied record identifies Kimi, Gemma, Qwen, DeepSeek, GLM, gpt-oss, Llama, Mistral, Phi, embedding, vision, and text-to-speech model families. Studio can search, download, and run GGUF files, LoRA adapters, and safetensors. The README claims training and reinforcement-learning support for more than 500 models. Compatibility varies by model format, operating system, hardware backend, and workflow; a complete compatibility matrix is available through the project documentation rather than the supplied metadata.
Observações sobre licença e riscos
The repository metadata declares the Apache-2.0 SPDX license. Detailed dependency, model-weight, dataset, and third-party service license terms are not stated in the repository metadata and may differ from the repository's own license.
Lançamento e manutenção
The supplied README highlights AMD training support, expanded GGUF hardware controls, local-agent connections, an MCP control endpoint, inference reliability work, newer model integrations, embedding fine-tuning, longer-context reinforcement learning, Triton kernels, packing improvements, and FP8 or vision reinforcement learning. Formal version numbers and release dates are not stated in the supplied repository record.