GitHub radar
An AI that answers with a number instead of a paragraph, instantly
This is a small model that skips the chatbot reply and gives you a calibrated answer right away — a number, a date, or a spot on a photo. Pair it with your usual chatbot, and routine decisions and data tagging get done in a fraction of the time.
It's a decision-making AI model: it answers yes/no questions with a probability, pulls numbers, dates, or text out of a document or photo, and finds objects in pictures — without generating any text.
When it helps
Useful when you need to process many simple decisions or extract data from text and photos, fast and cheap. Not a fit when you need a reasoned, written-out answer — that's a job for a regular chatbot.
Pros
- Answers in milliseconds — 5.7 ms on a server GPU
- Beats the larger d1-3B model while being smaller
- Understands 16 languages, including Russian
- Small enough to run on a laptop with 8+ GB RAM
Cons
- Doesn't generate text, so it can't hold a conversation
- Reliable on counts and scores, less so on open-ended estimates
- Speed was only tested on server GPUs, not laptops
How to set it up — step by step
- 1Open a coding AI agent (like Claude Code) inside your project folder.
- 2Send the agent the text from the block below.
- 3Wait while the agent installs the model and runs it on your machine.
- 4Give the agent your own sample text or photo to test it on.
- 5Check that the model's answer — a number or date — matches what you expected.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the d1-2B model from LiquidAI for me, from Hugging Face: https://huggingface.co/LiquidAI/d1-2B — run it locally and show me the calibrated answer it gives on my own text or photo example.
A small model, 1.92 billion parameters — by size this is laptop-class, 8 GB of RAM or more (this is an estimate based on the model's size), though the authors only publish speed numbers on server-grade H100 and MI325X GPUs.
Open on Hugging Face▌ More finds