Fine-tuneany LLMin one command_
Stop wrestling with training scripts. Soup gives you 11 training methods, 30 ready-made recipes, and 2-5x faster training with Unsloth — all from a single YAML config.
Integrates with your entire ML stack
HuggingFace
Ollama
vLLM
DeepSpeed
Unsloth
ONNX
NVIDIA TensorRT
W&B
SGLang
FlashAttentionHuggingFaceOllamavLLMDeepSpeedUnslothONNXNVIDIA TensorRTW&BSGLangFlashAttentionPerformance that pays for itself
Spend less on GPU hours. Train faster. Ship sooner.
Everything you need. Nothing you don't.
One CLI replaces your entire fine-tuning stack. No scripts, no boilerplate, no 200-line configs.
11 Training Methods
SFT, DPO, GRPO, PPO, KTO, ORPO, SimPO, IPO, Pretrain, Embedding, and Reward Model — all from a single CLI.
Vision + Audio Multimodal
Fine-tune models with text, image, and audio data. Full multimodal support out of the box.
One YAML Config
Configure everything — model, dataset, training params — in a single, readable YAML file.
30 Ready-Made Recipes
Pre-built configs for Llama 3.2, Qwen 3, Gemma 3, Phi-4, DeepSeek R1, Mistral — search, preview, and use instantly.
Unsloth 2-5x Speedup
Built on the Unsloth backend with Liger Kernel, FlashAttention, curriculum learning, freeze training, and loss watchdog.
Export Anywhere
Export to GGUF, ONNX, TensorRT, AWQ, GPTQ — deploy to Ollama, serve with vLLM/SGLang, or migrate from competitors in one command.
Three commands. That's it.
No boilerplate. No setup guides. No “just follow these 47 steps”. Install, init, train. Done.
You're spending too much time on infrastructure
Other tools make you fight the plumbing. Soup lets you focus on what actually matters — your model and your data.
Every hour spent on training infrastructure is an hour not spent improving your model.
Reclaim Your TimeAlready using another tool? Switch in 30 seconds
One command converts your existing config. No rewriting, no guessing — just migrate and train.
$ soup migrate --from llamafactory config.yaml
$ soup migrate --from axolotl config.yml
$ soup migrate --from unsloth notebook.ipynbSee the difference
model_name_or_path: meta-llama/Llama-3.1-8B
stage: sft
finetuning_type: lora
lora_rank: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target: all
dataset: alpaca_en
template: llama3
cutoff_len: 2048
per_device_train_batch_size: 4
gradient_accumulation_steps: 4
num_train_epochs: 3
learning_rate: 2.0e-5
lr_scheduler_type: cosine
warmup_ratio: 0.1
quantization_bit: 4
output_dir: ./saves/llama3-lora
base: meta-llama/Llama-3.1-8B
task: sft
data:
train: ./data/alpaca_en.jsonl
max_length: 2048
training:
epochs: 3
lr: 2e-5
quantization: 4bit
lora:
r: 64
alpha: 16
output: ./saves/llama3-lora
Soup auto-detects everything else — optimizer, scheduler, target modules, batch size.
Full Pipeline, One Tool
From training to deployment — Soup covers the entire LLM workflow.
Train
Fine-tune with 11 methods
Chat
Test your model interactively
Eval
Benchmark performance
Export
GGUF, ONNX, TensorRT, AWQ, GPTQ
Serve
vLLM, SGLang, transformers
Push
Upload to HuggingFace Hub
From pip install to deployed model in under 5 minutes
Start Building NowBuilt for the ML Stack you already use
First-class integrations with the tools powering production ML. Deploy anywhere, track everything.
Deploy & Serve
Training & Infra
Ecosystem
Export Formats
Works with your favorite models
Fine-tune any Hugging Face-compatible model. Soup supports all major architectures out of the box.
...and any model on Hugging Face Hub
Your competitors are
already fine-tuning.
Every day without custom models is a day your product falls behind. Soup gets you from zero to fine-tuned in under 5 minutes.