Story
A year ahead of the hype.
The method was published, open sourced and given away before anybody was selling it. Here is the whole line, from the first paper to the product.
The objective itself, written out →
2025
The papers came first
In March 2025 the first paper went up: reinforcement learning over sequence representations, with open weights, an open dataset and a package on PyPI. A second followed in September on schema based decisions guided by reinforcement learning.
Neither was a product. They were the argument that not every AI problem needs an autoregressive chatbot, made in the only form that can be checked, which is a method somebody else can run.
September 18, 2026
laya
A funded lab shipped the same idea as a closed, paid API, with no paper and no weights. A year of research was assembled into a release in a matter of days and open sourced under Apache 2.0.
Six thousand stars in three days, the front page of Hacker News, and within hours the community had given it a name nobody chose: the open decision model. It was ported to Node.js over ONNX Runtime, listed in model catalogues, offered on inference platforms and cited in a paper on pentest agents.
laya is community owned. It is not part of anything sold here and it never will be. The link below goes to the repository, not to a pricing page.
October 2026
vega, then vega-U
vega was the second family: small adapters trained on top of frozen backbones, in two sizes, with a benchmark harness pinned so anybody could rerun the comparison rather than take the table on faith.
vega-U went further and made the engine universal. One 14.3M decision engine in a shared space, and any frozen reader from 0.8B to 27B attaches to it through a small linear port. Readers it never trained with score highest, which is the result that says the engine learned the task rather than the reader.
Now
Embedding Studio
Every one of those releases produced the same question in the inbox, and it was never about the model. It was: how do I know I can trust this on my data.
That question does not have a research answer. It has a product answer, which is that you fit the thing to cases your own team already decided, you measure it on rows nothing was fitted on, and you attach a promise to the answer that somebody can hold you to. When no answer clears the bar, it returns nothing rather than guessing.
Embedding Studio is that, behind an API. It is the commercial product, it is what funds the research, and the engine underneath it is still open source.
The same thing, dated
March 2025
The first paper: reinforcement learning over sequence representations. Open weights, an open dataset, a PyPI package. A year before the hype.
September 2025
The second paper: schema based decisions guided by reinforcement learning.
September 2026
A funded lab ships the same idea as a closed, paid API. No paper, no weights.
September 18, 2026
laya. A year of research, assembled in days, open sourced under Apache 2.0. Six thousand stars in three days and the front page of Hacker News.
October 2026
vega, the second family. Two sizes of adapters on frozen backbones, with a benchmark harness anyone can rerun.
October 2026
vega-U, the universal adapter. One engine, any frozen reader, from 0.8B to 27B, attached through a small linear port.
Now
Embedding Studio. The research, turned into something a team can buy: your own decided cases in, a calibrated decision behind an API out, with a measured promise attached to every answer.