GitHub radar
Instant typed answers for text in 100+ languages, no waiting
This is an engine for fast typed decisions on text: instead of generating an answer word by word, it produces a verdict in one pass — a choice, a score, or a yes/no. Your app gets a result in milliseconds instead of waiting for a model to write one out.
A model that classifies text — picking an option from a set, scoring it, or answering yes/no — in one pass, without generating text step by step.
When it helps
Useful when you need to sort lots of text into ready-made categories (tickets, reviews, emails) quickly and cheaply, across many languages; not a fit for complex reasoning or tasks with many possible categories, where accuracy drops and a full language model works better.
Pros
- Answers almost instantly — about 33 ms on one GPU
- Runs without a GPU too — models are only 320–420M parameters
- Covers English plus 100+ other languages with auto-detection
- Code and weights are open under Apache 2.0, so you can self-host
Cons
- Accuracy on your own business tasks is just 0.362 without fine-tuning
- Struggles with many choices — only 0.425 accuracy on 77 options
How to set it up — step by step
- 1Install Laya with pip install laya in your project.
- 2Open the README at github.com/NandhaKishorM/laya and read the Router quickstart example.
- 3Send the agent the text from the block below.
- 4Check which model (English or multilingual) and what confidence the agent reported for each phrase.
- 5If accuracy on your own data is low, fine-tune the model on another checkpoint as described in the README.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up Laya from https://github.com/NandhaKishorM/laya following its README (pip install laya), run an English sentence and a non-English one through the built-in Router, and show me which checkpoint it picked for each and how confident it was.
Runs fine without a GPU: the models are about 320-420 million parameters and work on a regular CPU, while a basic T4 GPU or Apple Silicon answers one question in about 33 milliseconds.
Open on GitHub▌ More finds