Soup CLI

Soup CLI is a free, open-source command-line tool that handles the entire post-training pipeline for language models — from data preparation to training, evaluation, and deployment. It doesn't just launch a training job; it inspects your data, picks the right method (SFT, DPO, ORPO, SimPO, KTO, and more), writes the training config for you, derives eval tasks from your own data, and gates every saved checkpoint on a pass/fail verdict. With 23 methods, 142 recipes, and 17 quantization formats, it's a remarkably comprehensive toolkit for anyone fine-tuning or aligning open-source LLMs.

What makes Soup stand out is its layer streaming feature. Most fine-tuning tools require you to load the entire frozen base model into VRAM, which limits what you can train on consumer hardware. Soup keeps the frozen base in CPU RAM or NVMe and streams one decoder layer at a time into VRAM, quantizing it to 4-bit as it goes. On a 4 GB RTX 3050 laptop GPU, that means you can fine-tune Llama-3.1-8B at 119 tokens per second using only 3.32 GB of VRAM. The same streaming approach works for preference tuning methods like DPO, where the reference model is just the streamed base with its adapters turned off — no extra weights needed. Soup also includes a pre-flight check that refuses to start a run that won't fit your hardware, saving you from wasting GPU hours on doomed attempts.

Soup is built for machine learning engineers, AI researchers, and advanced hobbyists who want to fine-tune or align open-source models on their own hardware — especially when that hardware is a laptop or a single consumer GPU. It's ideal for experimenting with different training methods, running preference tuning without a server farm, or migrating from tools like LLaMA-Factory or Axolotl. The CLI is straightforward: initialize a project, point it at your data, and run. If you're already using another fine-tuning framework, a single `soup migrate` command converts your existing config. Soup Zero, a desktop workbench with a GUI and monitoring layer, is also in development.