• Picks
  • Discover
  • Alternatives
  • About

© 2026 Willet

  • Newsletter
  • About
  • FAQ
  • Terms
  • Privacy & Cookies
Join the Discord (opens in a new tab)
Alpamys Makazhan avatar

Soup

YAML-driven CLI for fine-tuning LLMs on consumer GPUs

  • AI models and runtimes
View on GitHub
Website

What it is

Soup is a command-line tool that fine-tunes and post-trains large language models from a single YAML config, covering SFT, DPO, GRPO and other methods with LoRA and QLoRA. It detects the GPU, picks batch size and quantization, and offers an opt-in layer streaming mode that keeps the frozen base model out of VRAM so an 8B model can train on a 4 GB card. Developers and hobbyists who want to adapt open models on their own hardware use it to skip cloud setup, and it can export results such as GGUF files for Ollama. A web UI is also included.

Images

Facts

Stars
9,460 GitHub stars.
9,460

Forks
1,555
Commits
1,957
Contributors
115
Language
Python
License
Apache-2.0
Last commit
Oct 20, 2026
On GitHub since
February 2026
First release
March 2026
Platforms
macOS, Windows, Linux, CLI
Self-hostable
No
Author
Alpamys Makazhan
Tags
  • model-training
  • reinforcement-learning
  • machine-learning
  • local-ai
  • fine-tuning

Last synced from GitHub on Oct 20, 2026

The 10 best overlooked open-source projects, every Monday.

Share and more

Comments

No comments yet. Start the conversation.

Similar projects

  • jingyaogong avatar

    MiniMind

    Train a tiny language model from scratch on a single consumer GPU

    64k
  • Unsloth AI avatar

    Unsloth

    Desktop and web app for running and fine-tuning local LLMs and diffusion models

    78.3k
  • Alex Jones avatar

    llmfit

    Terminal tool that matches local LLMs to your CPU, RAM, and GPU

    38.2k