Laya · open-source System One decisions
Convai's Laya is a 421M-parameter, Apache-2.0 non-autoregressive decision model: state + typed questions in, typed and calibrated decisions out, in one forward pass. It runs on your laptop. It's what the whole "typed decisions instead of text" idea looks like when it's open, and it's what I'd reach for first. Click any block.
Why Laya first
TypeSafe's Jev launched the "System One model" category and made the case beautifully — but it's closed weights, hosted API only, and the internal architecture is unpublished. Four days later, Convai shipped Laya: same category, same RLCD training idea, but under Apache 2.0 with a documented architecture built on ModernBERT. That's the version I want people building on.
The architectural shortcut
The interesting move is that Laya doesn't invent a new backbone at all. It reuses ModernBERT, which is a bidirectional encoder — parallel sampling comes for free because every token in the sequence already attends to every other token in one forward pass. There's no masking trick, no cache-reuse plumbing, no isolated branches. The state and the typed questions get packed into one sequence, the encoder does its usual thing, and a small 2-layer decision head on top turns embeddings into calibrated typed outputs.
Click the ModernBERT encoder block above to see what its attention mask looks like — it's the mechanism-level contrast to Jev's clever branch isolation.
What RLCD is doing
Only the decision head, the option-marker scorer, and the act/escalate head see gradients. The encoder stays frozen. That's why fine-tuning Laya on your own domain is cheap: proper scoring rules for the reward, GRPO-style updates, and the whole training loop fits in a public notebook. Calibration is the training objective, not an afterthought.
Where Jev still wins
Jev is faster on high-cardinality Choice tasks (50+ options in a single question) and its soft distributions look slightly sharper on some benchmarks. If you're building something where those matter and the closed hosted-only tradeoff is fine, it's the right pick — see the Jev walkthrough for that architecture.
Sources
- convaiinnovations/laya — the English 421M checkpoint on Hugging Face.
- convaiinnovations/laya-typed-decisions — the typed-decisions variant.
- github.com/NandhaKishorM/laya — reference implementation, RLCD fine-tune notebook, ONNX exports.
- answerdotai/ModernBERT-large — the backbone.
- receptron/laya — Node.js / TypeScript SDK via ONNX Runtime.