Soup

Soup is a local LLM fine-tuning tool for developers and teams adapting models to their own datasets. It brings training, evaluation and export into one workflow that runs on your own hardware without requiring a cloud service.

• Training methods: Use supervised fine-tuning, preference training such as DPO, or reinforcement learning with GRPO and PPO; QLoRA reduces the GPU memory needed to adapt a model.
• Local dashboard: Manage experiments, set up training, watch live metrics, explore datasets and chat with models through a browser interface served locally.
• Dataset support: Load chat conversations, preference pairs, plaintext, vision or audio data, with automatic format detection for JSONL, JSON, CSV, Parquet and TXT files.
• Model compatibility: Train Hugging Face text-generation models that load through AutoModelForCausalLM, including Llama, Qwen, Gemma, Mistral and DeepSeek.
• Model export: Export trained models to formats including ONNX, TensorRT, AWQ and GPTQ for use outside the training workflow.

Soup is free and open source under Apache 2.0, and supports Docker. It recommends a CUDA GPU and also supports Apple Silicon through MPS, with an optional MLX backend. Its hardware guidance calls for at least 8 GB of VRAM for roughly 7B models with QLoRA. Telemetry is off by default and requires explicit opt-in.