← Back to feed

ConvAI Innovations has released Laya, an open-weight model positioned as a direct alternative to TypeSafe AI’s Jev. The model runs at 421 million parameters on a ModernBERT backbone, matches Jev’s performance on structured classification tasks, and posts a median latency of 38.4ms compared to Jev’s 400ms average. It is available under an open Apache 2.0 license.

Laya reproduces Jev’s three-class output schema — Choice, Noul, and Score — making it a functional drop-in for systems built around Jev’s non-autoregressive decision architecture. Published benchmarks report Laya at 83.8% macro accuracy versus Jev’s 67.8% on structured decision tasks.

What Jev Is

TypeSafe AI launched Jev in September 2026 with a $40 million seed round. The model is non-autoregressive and designed specifically for structured decisions rather than open-ended generation. Rather than producing free text, Jev evaluates a choice space and returns one of three structured labels: Choice (select between options), Noul (boolean), or Score (continuous rating).

The architecture makes it fast and deterministic for decision support workflows where generative output introduces unnecessary variance. TypeSafe has not published weights, and that created immediate demand for an open equivalent.

What Laya Is

The 421M parameter count puts Laya in a range that runs on consumer hardware without quantization. ModernBERT’s encoder architecture is a natural fit: it processes the full input simultaneously rather than token by token, which gives it a structural latency advantage. At 38.4ms median versus Jev’s ~400ms average, the throughput difference is meaningful at production call volumes.

The speed gap is plausible even without seeing TypeSafe’s implementation. If Jev uses any form of iterative refinement or scoring passes, a direct encoder forward pass at 421M parameters would have a structural throughput advantage regardless of absolute model size.

Matching Jev’s three-class output schema is the practical contribution. Prior reproduction attempts — over a dozen tracked by the community-run Jev Reproductions Tracker as of mid-September — used available open models to approximate Jev’s behavior. Laya is among the strongest contenders, publishing weights alongside benchmark numbers that exceed Jev’s reported accuracy.

What This Does to Jev’s Position

TypeSafe’s commercial case for Jev rests on three potential differentiators now that Laya covers the core use case with open weights: proprietary training data, compliance certification for enterprise deployments, and customer support.

If Laya performance parity holds under independent evaluation, the model’s open nature removes the weight-access barrier and the latency disadvantage simultaneously. Teams building Jev-based pipelines have a credible open baseline to replace or supplement the TypeSafe API.

TypeSafe’s response will determine how durable that differentiation is. The $40M seed was raised on the premise that structured decision AI is a distinct product category. Laya does not invalidate that premise — it applies competitive pressure to the closed-weights implementation of it.