Soup

Soup is an open source command line tool for post training large language models on local hardware, built by MePlay, Inc. and released under the Apache 2.0 license. It installs with pip and covers the workflow from data checks to shipping a model in a single CLI. Its central technique is layer streaming. Instead of loading the whole frozen base model into GPU memory, Soup stores it as one shard per decoder layer in system RAM or on NVMe storage and copies layers into the GPU one at a time while the next one loads, so only the LoRA adapters stay resident. With optional 4 bit quantization of the streamed base, the vendor reports that an 8B Llama model can be fine tuned on a laptop GPU with 4 GB of memory, and that preference methods such as DPO, ORPO, SimPO and KTO stream as well. Beyond that, Soup checks datasets before training, picks a method and writes the configuration from rules rather than a search, derives evaluations from the user's own data, gates checkpoint saves, and responds to reward hacking during a run. The site lists 23 training methods, 167 recipes and 17 quantization formats. A migrate command converts existing configurations from LLaMA Factory, Axolotl and Unsloth, and it integrates with Hugging Face, Ollama, vLLM, DeepSpeed, ONNX, TensorRT and Weights and Biases. It is aimed at practitioners who want to fine tune models without renting larger GPUs. The tool is free. A desktop workbench called Soup Zero is announced as coming soon.