I Found an AI That Doesn’t Want to Talk. It Just Decides.

Most AI today is judged by how well it talks.

We ask it to explain things, write things, summarize things, reason out loud and turn almost every problem into a conversation.

Laya takes a strangely different route. It does not try to answer in prose at all.

You give it some text or structured data and a few questions: which department should handle this email, how urgent is it, does it look like phishing, which option fits best. Laya returns a choice, a score or a probability.

That sounds almost like a step backwards. After all, we have spent years making models more general, more articulate and more capable of doing everything.

But a surprising amount of work inside software is not really a conversation. It is a decision.

Route this request. Flag that message. Pick one of four actions. Decide whether a human should look at something. Estimate how strongly a piece of text matches a category.

Using a large language model for those jobs can mean asking a very powerful machine to generate an invisible little answer and then hoping the surrounding software parses it correctly.

Laya skips the talking part.

The project is open under Apache 2.0 and ships several checkpoints. The main English model is about 421 million parameters; a multilingual version is about 322 million and covers more than 100 languages. They can run on your own hardware rather than behind a mandatory cloud API.

More importantly, the model does not generate output tokens one after another. It scores the possible decisions directly in a single forward pass.

There is an important catch. Laya is not a tiny universal judge that can magically replace a larger model. Its own published results show that the general checkpoints can be weak on tasks they were not specialised for, and the documentation repeatedly tells users to evaluate and fine-tune for their own workload.

I actually like that limitation.

The interesting idea is not that a 322-million-parameter model has somehow beaten every chatbot. It is that some AI jobs may never have needed a chatbot in the first place.

We keep pushing intelligence toward one enormous general system that can talk about anything. Laya points in the other direction: smaller systems with a very specific job, running locally, returning only the kind of answer the software actually needs.

It will not replace ChatGPT, Claude or Gemini. That is not the point.

Sometimes the right AI is the one that has almost nothing to say.