Fine-tune LLMs from one YAML file. Soup uses layer streaming so an 8B model can be trained on a 4 GB laptop GPU, and covers DPO as well as supervised fine-tuning with GGUF export.
Be the first to rate Soup. Your feedback helps other users decide.
Train an 8B model within 4 GB of VRAM using layer streaming.
Describe the whole run in a single YAML instead of a training script.
Produce GGUF weights you can run in a local inference server.
Similar tools in the same category β compare side-by-side
A local AI-powered red teamer that runs on a phone. Nightcrawler brings an agent-driven security tes...
Zero-cost epistemic uncertainty quantification and hallucination detection for LLMs. Spanda estimate...
Compile-time LLM code synthesis with deterministic, self-routing NLU runtimes. Strot compiles an LLM...
Guides and comparisons about code tools
Eleven recent open-source projects for running inference, transcription, deep research and security work entirely on hardware you control.
Ten Show HN projects for running several agents at once: orchestrators, subagents in isolated worktrees, shared memory trees and browser relays.
Eight Show HN projects that handle the unglamorous layer: consumer-GPU serving, remote tensors, vector-ready storage, semantic layers and agent payments.
Which AI coding assistant actually saves you time? GitHub Copilot vs Cursor vs Codium in 2026, with real comparison and recommendations by use case.
How to choose an AI coding tool in 2026: inline completion, repo chat, and local privacy, compared across Copilot, Cursor, Codeium, and more.