Laya vs Jev
You found Laya — the open-weight decision model — and now you want to know how it stacks up against Jev, the hosted original. Both answer typed questions instead of writing text; the split is open-and-local versus hosted-and-calibrated. Here's the honest comparison, including what Laya's own benchmark claims do and don't prove.
Laya (Convai Innovations) shipped open weights under Apache 2.0 a few days after TypeSafe launched Jev, and it targets the exact same job: a state and typed questions in, a decision with a probability distribution out — no free text. The difference is ownership. Jev keeps the model hosted and calibrated; Laya hands you the weights to run, fine-tune and pin yourself.
Laya vs Jev at a glance
| Laya (open) | Jev (TypeSafe) | |
|---|---|---|
| What it is | Open-weight System One model | Hosted System One API |
| Weights | Open (Apache 2.0) | Closed, versioned & pinned |
| Runs | Local — MLX (Mac), ONNX (Node), GPU | Hosted — just a key |
| Calibration | Your responsibility; validate it | RLCD-trained, calibrated |
| Latency (reported) | ~33 ms (model card); ~7–14 ms MLX in-process | ~70–500 ms incl. network |
| Cost | Free to run; you pay compute | ~$0.001 / decision, output free |
| Setup | Install weights + runtime | Minutes — grab a key |
| Best for | Privacy, offline, fine-tuning, volume | Production, no ops, trusted numbers |
About Laya's benchmark claims
Laya's model card claims it beats Jev — figures like ~83.8% vs 67.8% accuracy and ~33 ms vs ~236–276 ms latency circulate widely. Read them as vendor-reported on Laya's own benchmark, not a neutral head-to-head. Community game demos that reward low latency flatter fast local models, and calibration — whether a stated 0.8 is right ~80% of the time — is exactly what an open checkpoint doesn't guarantee. The honest takeaway: Laya is impressively fast and clearly capable, but validate accuracy and calibration on your own labelled data before you trust a claim.
When Laya wins
Choose Laya when control and economics dominate: the decision must stay on your hardware (regulated data, offline), you want to fine-tune on your domain, or you're running such high volume that per-call pricing hurts. Its MLX build runs in-process on a Mac with no server and no bill, which is genuinely hard to beat for local dev and privacy.
When Jev wins
Choose Jev when you want a calibrated number you can threshold and forget. Because the weights are pinned and RLCD-trained, a rule like "escalate if p > 0.8" keeps meaning the same thing between deploys instead of drifting, and you carry no GPUs, no weight updates, no calibration validation. For production decisions where a wrong-but-confident answer is expensive, that reliability is the whole point.
Run both — or try each
Many teams will pair them: Laya for local, private, high-volume bulk work and Jev for the calls that need a pinned, calibrated contract. Both speak the same primitives, so your interface stays the same while you swap the backend. Want to deploy Laya? The step-by-step is on the Laya local guide. Want to feel the hosted side? The playground runs a real calibrated decision in the browser, free.
See also: Deploy Laya locally (guide) · Jev vs Laya (hosted-first view) · NanoJev vs Jev · Playground
Try the hosted, calibrated side free
Before you commit to self-hosting Laya, run a real calibrated decision in the browser — no signup — then grab a jv_live_ key if hosted fits.