In March 2025 I published a paper (arXiv:2503.23303) on a model that was not autoregressive, did not generate text, and returned probability trajectories over structured outcomes, trained with reinforcement learning over sequence representations. The weights went on Hugging Face, the dataset went public, the package went on PyPI, and the idea went on Reddit. In September 2025 a second paper (arXiv:2510.01237) formalized the framework: schema-based decisions guided by reinforcement learning.
A year later, in September 2026, a well-funded lab shipped the same idea as a closed, metered API and the internet called it a breakthrough. No paper, no weights, no training data.
The frustration was real, and so was the choice it forced: stay bitter or build the open version properly. laya is the second option.
What a decision model is
Most of what software asks a model is not an essay. Which queue owns this ticket. Is this email phishing. Is this prompt an injection attempt. How urgent, on a rubric from 0 to 3. Calling a generative LLM for these means waiting for tokens, parsing free-form text back into a label, and trusting a confidence number that is itself just a plausible-sounding token.
A decision model answers typed questions, choice, score, or yes and no, in a single forward pass, and returns probabilities that mean something because the training loss was a strictly proper scoring rule. The output space is numbers, so hallucination and malformed JSON are not failure modes; they are impossible by construction.
What a year of head start bought
- Speed: about 33 ms per decision on a T4, 7.2 ms per question batched, because a 421M bidirectional encoder is the right size for a reflex.
- Honest probabilities: expected calibration error of 0.081 after temperature fitting, against 0.246 published for the hosted alternative.
- Languages: a multilingual sibling measured across 51 languages, with a sub-millisecond router that picks the right checkpoint by script before the forward pass, because confidence cannot warn you about a tokenizer that shreds the input.
- Openness: Apache 2.0 weights, training code, and every benchmark with the harness that produced it.
The launch went to the front page of Hacker News, 1,363 points, and collected six thousand GitHub stars in three days. The more durable outcome is the ecosystem: ports to Node.js, hosted offerings, write ups, and a citation as a System One decision layer in other people's research.
Honesty about the ceilings
laya degrades past 20 options in a choice question, scores near random zero-shot on benchmarks it was not tuned for, and ships with raw temperatures you should fit on your own distribution. All of that is in the README, because a model you can only trust when its authors are in the room is not infrastructure.
The full story, with every number and link, is in the original dev.to piece. Where it goes next, vega and the universal adapter, is the rest of this blog. And laya itself stays where it belongs: community-owned, open for good.