GitHub radar
Filter simple yes/no questions cheaply before calling a big model
d1-3B is a small decision model: it doesn't write text, it instantly returns a probability for a yes/no question, a choice, or a rating. Use it to handle simple cases yourself instead of running every request through a big model.
Liquid AI builds small models meant to run on everyday devices — the family is called LFM. d1-3B is their decision model, built with 3 billion parameters. It's based on LFM2.5-VL-3B, a model that reads both text and images.
When it helps
Use it to cheaply filter simple cases — spam, ticket routing, photo moderation checks — before sending the hard ones to a bigger model. Skip it if you need normal text replies: this model can't write text.
Pros
- Beats every model under 10B params on the maker's own benchmark.
- Answers in milliseconds, even on a laptop or phone.
- Handles text and images together in one request.
- Try it in the browser with no install.
Cons
- Most numbers are the maker's own tests, not independent benchmarks.
- Can't write text or hold a conversation — only yes/no/score answers.
How to set it up — step by step
- 1Open the Liquid AI d1 Arcade demo in your browser: huggingface.co/spaces/LiquidAI/system-one-arcade.
- 2Test your own question on the demo — yes/no, a choice, or a rating — and check the probability it returns.
- 3To use the model in your own project, open an AI agent (Claude Code, Codex) in your project folder.
- 4Send the agent the text from the block below so it checks the model on your example.
- 5Check the agent's reply: see whether the probability the model returns fits your use case.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the d1-3B demo with no install: https://huggingface.co/spaces/LiquidAI/system-one-arcade — try it on my example (a yes/no question or a pick from options) and show me what probability the model returns.
The model has 3 billion parameters — it fits on a regular laptop with 8 GB of RAM or more, and answers in a few milliseconds on a GPU or an Apple M-series chip.
Open on Hugging Face▌ More finds