On this page (7 sections)
Short answer: Laya is the better Jev alternative if you want calibrated probabilities, Swiss data residency with zero retention, and a drop-in replacement for TypeSafe Jev's /v1/systemone protocol. Jev still holds the accuracy edge, scoring 0.870 to Laya's 0.425 on Banking77 in our published benchmarks. Laya's open model has 5.85k likes on Hugging Face, and our hosted API answers typed questions in a single forward pass, typically in about 120 ms from Europe.
Key takeaways
- Laya is a hosted decision API, not a text generator: it answers typed questions (choice, score, yes/no) with a calibrated probability for every question and cannot hallucinate a label.
- It is compatible with TypeSafe Jev's /v1/systemone protocol, so you can repoint existing Jev requests without rewriting your schema.
- Pricing is 100,000 free questions per month, then $5 per million, with no hidden tiers.
- Laya runs in Switzerland with zero data retention and supports English plus 100+ languages.
- The honest trade-off: Jev leads on raw Banking77 accuracy; Laya leads on residency, retention, calibration, and predictable spend.
What is TypeSafe Jev, and how does Laya relate to it?
TypeSafe Jev is the decision platform that originated the /v1/systemone protocol. Laya is an independent, Swiss-hosted decision API built to speak that same protocol, so questions you already send to Jev—choice, score, or yes/no—can be repointed with little schema change. We are not affiliated with TypeSafe or Convai Innovations; the open-source Laya model is on Hugging Face.
As of October 2026, Jev's own published pricing, retention, and hosting terms are not available in the sources I can verify, so I won't quote them here. What I can say is that Laya's compatibility claim is documented on Laya Studio, and the open model itself is listed at convaiinnovations/laya on Hugging Face, which shows 5.85k community likes.
What is a decision API?
A decision API takes text or JSON plus typed questions and returns structured answers instead of free-form text. With Laya, you send a message and a question such as "Is this a refund request?" and get back a calibrated probability, not a generated sentence. Because no text is produced, the model cannot hallucinate a label or invent a category that does not exist. That makes it suitable for automated routing, triage, moderation, and escalation where the output must feed directly into another system.
What are calibrated probabilities?
A calibrated probability is a number that reflects how often the model is right: a score of 0.80 means the answer should be correct about 80% of the time. Laya returns one probability for every question in a single forward pass. This matters operationally because you can set hard thresholds—escalate to a human only when confidence falls below 0.50, for example—rather than parsing ambiguous strings from an LLM prompt. Calibration also makes failures auditable; you can see exactly where the model was over- or under-confident.
Is Laya a Swiss AI API?
Yes. Laya Studio runs in Switzerland with zero data retention, and it supports English plus 100+ languages. As of October 2026, that means classification traffic is processed under Swiss data protection rules and is not stored after the response is returned. For health-tech teams and support desks handling personal data, that residency and retention posture is a concrete compliance difference, not a marketing claim.
How do the two APIs compare?
The table below contrasts the dimensions that matter most for a routing or triage pipeline. Where a detail is not published in the sources I can verify, I have said so rather than guessed. All Laya figures are as of October 2026 and come from Laya Studio; the benchmark scores come from our published results on GitHub.
| Dimension | Laya | TypeSafe Jev |
|---|---|---|
| What it is | Hosted decision API for the open-source Laya model (Apache-2.0) | Decision platform behind the /v1/systemone protocol |
| Protocol | Compatible with Jev's /v1/systemone | Native /v1/systemone implementation |
| Output | Calibrated probability per question; no text generated | Classification labels via protocol |
| Can hallucinate labels | No | Not published in sources reviewed |
| Data residency | Switzerland | Not published in sources reviewed |
| Data retention | Zero retention | Not published in sources reviewed |
| Pricing | 100,000 free questions/month, then $5 per million | Not published in sources reviewed |
| Latency from Europe | ~120 ms per forward pass | Not published in sources reviewed |
| Languages | English + 100+ languages | Not published in sources reviewed |
| Banking77 score | 0.425 (our benchmarks) | 0.870 (our benchmarks) |
The honest trade-off runs both ways. Jev's genuine advantage is raw accuracy on published classification benchmarks—on Banking77 it scored 0.870 to Laya's 0.425—and as the originator of the /v1/systemone protocol, its client libraries are the reference implementation. If you are already running Jev and your only goal is maximum intent-classification accuracy, staying put may be the right call.
Laya's advantages are operational. You get calibrated probabilities instead of uncalibrated labels, Swiss hosting with zero data retention, transparent per-question pricing, and about 120 ms latency from Europe. Because it speaks Jev's protocol, migration is a repointing exercise, not a rewrite. For teams that need auditable decisions, data residency, or predictable spend, Laya is the better fit.
Who should choose Jev, and who should choose Laya?
Choose TypeSafe Jev if you are already deep in its ecosystem, you need the highest published accuracy on Banking77-style intent classification, and data residency or zero retention is not a requirement for your workload.
Choose Laya if you need a drop-in Jev replacement that returns calibrated probabilities, processes data in Switzerland with zero retention, supports 100+ languages, and charges predictively per question. It is especially suited to support routing, clinical message triage, content moderation, and any pipeline where the output must be consumed by code rather than read by a human.
If you want to try it, the first 100,000 questions are free on Laya Studio.
FAQ
Can I use Laya as a drop-in replacement for TypeSafe Jev?
Yes. Laya is compatible with TypeSafe Jev's /v1/systemone protocol, so existing requests can usually be repointed without changing the request schema.
Does Laya generate text or hallucinate labels?
No. Laya returns a calibrated probability for each typed question and generates no text, so it cannot invent a label or category.
Where is Laya hosted, and is my data retained?
Laya runs in Switzerland with zero data retention; traffic is processed and not stored after the response is returned.
What is a calibrated probability, and why does it matter?
It is a probability that reflects how often the model is correct, letting you set firm escalation thresholds instead of parsing free-text LLM output.
How much does Laya cost?
As of October 2026, Laya offers 100,000 free questions per month, then $5 per million questions.
Topics
- Jev alternative
- TypeSafe Jev
- decision API
- calibrated probabilities
- Swiss AI API