Soup CLI

Open-source CLI for fine-tuning LLMs with layer streaming, preference training, and eval-gated deployment.

Overview

Soup CLI is a comprehensive open-source command-line tool for fine-tuning large language models on consumer GPUs. It automates the entire post-training workflow—from data curation and training method selection through evaluation and deployment—with a focus on correctness and efficiency.

The tool's flagship feature is layer streaming, which allows models like Llama-3.1-8B to fine-tune on 4 GB GPUs by keeping the frozen base in CPU RAM or NVMe and loading decoder layers one at a time. Soup supports 23 training methods (SFT, DPO, ORPO, SimPO, KTO, GRPO, and others), 170 offline recipes, and multiple backends including Transformers, MLX (Apple Silicon), and Unsloth. It includes built-in data validation, semantic deduplication, preference training with verifiable rewards, and eval-gated saves that prevent shipping models that break general knowledge.

Soup is free, Apache-2.0 licensed, and designed to run entirely on your own hardware with no vendor lock-in.

Key features

Layer streaming for 4 GB GPU fine-tuning
23 training methods (SFT, DPO, ORPO, SimPO, KTO, GRPO)
170 offline recipes
Eval-gated saves and model verification
Semantic data deduplication
Preference training with verifiable rewards
Multi-GPU support (DeepSpeed ZeRO)
Apple Silicon (MLX) backend
GGUF export and quantization
MCP server for coding agents

Pros

Runs on consumer GPUs (4 GB VRAM with layer streaming)
Comprehensive post-training pipeline in one tool
Free and fully open-source (Apache-2.0)
No vendor lock-in; works offline
Extensive documentation and 170 ready-to-use recipes
Strong focus on correctness and reproducibility
Eval-gated deployment prevents shipping broken models
Active development with transparent release notes

Cons

Layer streaming is BETA and slower than resident training
Steep learning curve for advanced features
Requires Python 3.10–3.12 (not 3.13+)
Large dependency footprint for full install
Configuration complexity for non-standard setups

Use this if

You need to fine-tune LLMs on limited VRAM, want a complete post-training pipeline with eval gating, require reproducible and auditable training, or need to work entirely offline without vendor dependencies.

Skip this if

You need a managed cloud service, prefer graphical interfaces over CLI, require commercial support, or are using Python 3.13+.

Best for

Fine-tuning LLMs on limited VRAMPreference training and alignmentData validation and semantic deduplicationReproducible ML workflowsSupply-chain security and complianceMulti-GPU training with DeepSpeed

Alternatives

LLaMA-FactoryAxolotlUnslothTRL (Hugging Face)Ludwig

Compare Soup CLI alternatives

More AI

Compare all

coolplugz

Orchestration layer that guides Claude Code to deliver coding tasks without constant supervision.

Goldfish

AI assistant that remembers your work and writes replies in your tone, available for macOS and Windows.

Runtime

AI agent platform for payment teams to automate operations, risk, compliance, and support workflows in their own cloud.

Scribble AI

AI-powered blog writing tool that generates SEO keywords, outlines, first drafts, and hero images from a title.