Skip to content
Soup Zero · Coming soon · the deep layer is live today

Own your model.

The AI workbench for post-training, evals, and deployment. Cursor for people who work with AI models: one workspace to do everything with open models, on your own hardware.

Install the CLISee what's live today
Open source
Soup Zero · llama-3.1-8b · support-copilot
Fine-tune · run #42No-codeCopilotCode
Training llama-3.1-8b on support-tickets.jsonl
Helpfulness
Accuracy
General skills
Evals passed → SHIP · reproducible
Ask Soup to do it▋

Your data. Your GPU. Your model.

Concept preview. The engine underneath (fine-tuning, evals, deploy) ships today as the open-source Soup CLI.

You do not want to rent a model forever. You want to own one.

  • The easy tools run out.

    Click-through platforms stop at a wrapper. The first time you need a real post-training run, you are done with them.

  • The real ones are a pile.

    A notebook, four scripts, a dozen CLIs and a spreadsheet of results. Nothing checks whether the last run made the model worse.

  • Your data should not travel.

    Fine-tuning in someone else's cloud means your training set and your weights live on their disk. Often you simply cannot.

What is missing is one workbench that holds all of it.

One workbench for the whole lifecycle.

One desktop app, a station for every stage. Three of them ship today as the open-source Soup CLI; the rest are in development.

Live todayshipping in the Soup CLI, one command each

  • Fine-tuning

    Train it on your data.

    $ soup train --config soup.yamlFine-tuning docs

  • Evals

    Measure quality before you ship.

    $ soup eval gate --suite evals/gate.yaml --model ./outputEvals docs

  • Deploy

    Serve it from your machine behind an OpenAI-compatible API.

    $ soup serve --model ./outputDeploy docs
In developmentbuilding it now

  • Playground

    Try open models instantly.

  • Data

    Collect, clean and label your data.

  • Prompt Studio

    Write and compare prompts.

  • RAG / Knowledge

    Connect your documents to the model.

  • Agents & Tools

    Give your model tools (MCP).

  • Monitoring

    Watch quality in production.

  • Team

    Share and review together.

Your hardware. Your data. Your model. The deep layer (fine-tuning, evals, deploy) is live today as the open-source Soup CLI, and it runs on your own machine. The workbench around it is what we're building.

Click it. Say it. Or code it.

Every station, three ways in. Pick one per task, switch anytime.

Live today Full control in the terminal, the open-source Soup CLI.

Terminal

$ soup train --config soup.yaml

$ soup eval custom --model ./output --tasks quality.jsonl

$ soup ship --base Qwen/Qwen2.5-3B --adapter ./output --task-eval quality.jsonl --config soup.yaml

SHIP ▋

From raw data to your own AI model.

Upload data, train, check quality, deploy safely, improve over time. One loop.

  1. Open model
  2. Playground
  3. Data
  4. Fine-tuneLive today
  5. EvalsLive today
  6. DeployLive today
  7. Monitor
  8. ...and production feedback flows back into Data.
Live today$ soup ship --evidence evidence.json
SHIP

Every run ends with a verdict.

SHIP or DON'T SHIP, with the scores behind it in a file you can commit and replay to the same verdict.

  • Your task metric beat the base model
  • No general benchmark fell past the forgetting threshold
  • Exit code 0 on SHIP and 2 on DON'T SHIP, for CI
Soup Zero

The deep layer is live today.

Fine-tuning, evals and deploy are not a preview. They ship as the open-source Soup CLI, they run on your own hardware, and you can have them now. The workbench is what we are building around them.

Install the CLI

Built on the open-source Soup CLI · PyPI · team@trysoup.dev