Jev vs Laya

Jev vs Laya is less about which model is smarter and more about where your decision layer runs. Use TypeSafe's hosted Jev when you have no labelled data. Self-host Laya when latency or data residency matters more, and accept that you now run a model.

What is Jev?

Jev is TypeSafe AI's hosted decision model: you send state plus typed questions and get answers with probabilities, not text. TypeSafe's launch post opened early access on 15 September 2026, pricing input at $0.042 per million tokens with output free.

The Pydantic AI docs and jevwiki agree on context: 64k tokens per request, with 32k for the state plus the longest single question. Pydantic AI's docs cap a choice at 255 options and advise pinning jev-1.13.0 once thresholds are tuned, because jev-latest moves with each release.

The limitation is control: closed weights, no fine-tuning. laya.studio's comparison notes that TypeSafe doesn't publicly document where requests are processed, and that enterprise customers get zero data retention.

What is Laya?

Laya is an open-weight decision model from Convai Innovations that accepts the same POST /v1/systemone request as Jev. Wilson Wu's write-up (23 September 2026) records that Nandakishor M released it on 18 September under Apache 2.0, at 421M parameters, running on a GPU, a Mac or a CPU. laya.studio lists multilingual and fine-tuned typed-decisions checkpoints too.

The limitation is zero-shot quality. laya.studio reports the base checkpoint at 0.362 on the typed-decisions benchmark, below a 0.461 majority-class baseline, while the 0.766 headline needs the fine-tuned checkpoint. The model card says it plainly: Laya "is a fast base to specialise, not a zero-shot decision engine."

Jev vs Laya: quick comparison

• What it is: Jev is TypeSafe's closed, hosted API; Laya is an open-weight model under Apache 2.0.
• Primary strength: Jev is strong zero-shot on long inputs; Laya is fast and can be fine-tuned.
• Common uses: Jev for routing with many labels; Laya for guardrails and short, high-volume decisions.
• Where it runs: Jev runs in TypeSafe's cloud (location not publicly documented); Laya runs on your infrastructure.
• Context: Jev allows 32k for state plus the longest question (Pydantic AI docs); laya.studio gives Laya 512 tokens per question on the English checkpoint and 1,024 on the others.
• Options per choice: Jev takes up to 255 (Pydantic AI docs); laya.studio advises under about 20 for Laya, citing Banking77 scores of 0.870 for Jev and 0.425 for Laya.
• Latency: laya.studio cites 236 to 276 ms p50 end to end for Jev (independent benchmarks) and 32.8 to 39.5 ms on a T4 for Laya (model card).
• Cost: Jev is $0.042 per million input tokens (TypeSafe); Laya has no licence fee, so you pay for hardware and operations.

laya.studio warns the latency figures aren't like for like, since Jev's include the internet round trip. Published Jev vs Laya accuracy claims disagree: a Hugging Face community post (24 September) reports JevBench v1.3.0 ranking Jev first at 74.4 and untuned Laya 33rd at 54.4, while Laya's own tables favour it on short, few-label tasks. Shadow test on your own traffic.

What running Laya in production involves

When we put a model like this into production, we would start with the network. laya.studio notes that laya-serve binds to all interfaces with no authentication unless LAYA_API_KEY is set, so put it behind your own gateway and TLS, as you would when securing AI systems generally. Wilson Wu notes a language switch in lazy mode costs a 7 to 10 second reload, so preload.

ZimaSpace's comparison points out that owning the checkpoint lets you pin a model version and retest upgrades before production behaviour changes.

laya.studio, citing the model card, reports mean ECE (the gap between stated confidence and actual accuracy) falling from 0.466 as shipped to 0.081 after refitting temperatures on held-out data. Do that before thresholds drive actions. The two models compute confidence differently, so thresholds don't copy across. Then watch for drift, as in LLMOps in practice.

laya.studio notes Laya ignores Jev model IDs, so the same body can fall back to Jev for many-option or long-state questions.

Hosted API or self-hosted: which do you need?

ICO guidance treats sending personal data to a US processor as a restricted transfer. That needs the UK Extension to the EU-US Data Privacy Framework, which only covers recipients with an active status on the DPF list, or a safeguard such as the IDTA with a transfer risk assessment. Check the list, ask for a DPA and strip personal details first. Self-hosted Laya in a UK region avoids the transfer.

Cost rarely decides it. Wilson Wu puts Jev at about $42 per million decisions at 1,000 input tokens each. A T4 running all month at roughly $0.50 an hour costs about $360, the price of about 8.5 million Jev decisions, before redundancy and staff time.

My stance: start on Jev to prove the questions, collect reviewed labels, then move short, sensitive, high-volume decisions to a fine-tuned Laya you host, keeping Jev as fallback. Jev rents you a decision layer; Laya hands you one to run.

We scope AI integration work first and quote a fixed price per agreed release, with code and cloud accounts in your name. Whether it's Laya on your own cloud or Jev wired into your stack, talk to an engineer at Eight Mile.