Sovereignty · 2026-09-29 · 1:48

Yours — after you train it

Name: Yours — after you train it
Uploaded: 2026-09-29
Duration: 1 min 48 s
Description: The correction is to our own optimism. The primitive can be yours, and the price is a labelled set from your own work, a fine-tune, and a calibration refit before you trust a single probability. Owning the weights is the cheap part.

A week ago we wrote that decision models, which never write a sentence and only pick from the answers you allow, had become a primitive you can own. Days after the paid Jev appeared, an open alternative arrived: Laya, Apache 2.0, 421M parameters, weights in hand. That is still true. What we did not say is what owning it costs.

In favour

Genuinely yours: open weights, self-hosted, no per-token fee, and 32.8 ms per question on a modest GPU against 236 to 276 ms measured by third parties over the paid API.
The model card is unusually honest, and that honesty is the reason we can write this at all.

Worth watching

Out of the box it is near chance. On its own benchmark of 2,000 decisions the base English checkpoint scores 0.362, while always guessing the most common answer scores 0.461. The card's words: "a fast base to specialise, not a zero-shot decision engine."
Fine-tuned it reaches 0.766, but the card notes that checkpoint was trained on that benchmark's own training split. With 77 options it scores 0.425 against the paid model's 0.870. And it ships over-confident: on Khmer, a script the English checkpoint cannot read, 0.000 accuracy at 0.952 confidence. "Confidence gating cannot save you."

Our takeThe correction is to our own optimism. The primitive can be yours, and the price is a labelled set from your own work, a fine-tune, and a calibration refit before you trust a single probability. Owning the weights is the cheap part.

Source: Κάρτα μοντέλου Laya (Convai Innovations, Hugging Face)

← All Focus posts

AI-narrated · AI-generated content

Transcript

Last week we wrote that a decision model can be yours. That is still true. We didn't say what owning one costs. An open one arrived days after a paid one. Apache 2.0, on your own hardware. It never writes text. It only picks from the answers you allow. Its own model card is plain about it. Out of the box, it is "a fast base to specialise, not a zero-shot decision engine." On 2,000 decisions, the base model scores 36%. Always guessing the most common answer scores 46%. Fine-tuned, it reaches 77%, past the paid one's published 73%. Speed is 33 ms on a local GPU, against a quarter of a second over someone else's API. Not the same race, the card says. Worth watching. That fine-tuned model was trained on the same benchmark's own training split. And with 77 possible answers instead of a handful, it scores about half of the paid one. And it ships over-confident. On Khmer text, which its English version can't read, it got nothing right while reporting 95% confidence. "Confidence gating cannot save you." None of this makes it a bad choice, only a different one. Sovereignty is paid in data and effort: your own labelled examples, calibration and test set. As the short that pointed us here put it: test it on your own labels.