Laya, free for eighteen months, hit 17,900 stars in five days
Nandakishor Mukkunnoth's Apache-2.0 decision model took off only after TypeSafe shipped its paid Jev.
A model that had been sitting on Hugging Face for eighteen months went from zero to 17,900 GitHub stars in five days. It is called Laya, it is Apache 2.0, and the weights were public the entire time.
Nandakishor Mukkunnoth, at ConvAI Innovations, published the paper in March 2025. Non-autoregressive decision models trained with RL. Nobody paid attention. Then TypeSafe, founded by one of the ChatGPT co-inventors, shipped Jev behind a metered API, and he posted that he had built the same class of thing a year earlier and given it away for nothing. The post hit 1,338 points on Hacker News, and the stars arrived that week.
The mechanism is simple enough to explain in a sentence. 421 million parameters, one forward pass, a typed answer out the other side. It does not write anything. There is no token stream to parse, no malformed JSON to repair, no chance of it inventing a sixth option when you handed it five. It covers 45+ languages and lands decisions under 35 milliseconds. His measurements put Laya at 33 milliseconds against Jev's 236, which is where the 7.8x latency figure comes from, alongside a claim of roughly three times better calibration.
Calibration is the number I would actually check. Latency you can buy your way out of with better hardware. A confidence score that means what it says is what lets you set a threshold and route the uncertain cases somewhere else, and most of the classifiers people bolt onto agents today are confidently wrong in ways nobody measures.
The community got there fast. There are already MLX and Core ML ports doing a decision in 5 to 13 milliseconds on an M3 Max. That is a local, free, sub-10-millisecond answer for work most of us currently send over the network and pay per million tokens to get back.
Worth saying plainly: these are his benchmarks. Nobody independent has reproduced the head-to-head, and any comparison like this bends to whoever picks the task set. The difference is that Apache 2.0 weights let you run the test yourself, which you cannot do with Jev.
What I think is that a large share of what we call agent work is classification in a trench coat. Routing, yes-or-no gates, pick one of five, does this need a human. We pay frontier prices for those because the frontier model was already open in the tab and wiring up a second thing felt like effort. A 421M model that answers in one pass handles that tier, and the eighteen months of silence around Laya says the hard part was never building it.