Jev vs Laya: an honest comparison

TypeSafe Jev and Laya do the same job. Each reads a message or record, answers typed questions about it (pick one, rate, yes/no) with a probability for every option, and writes no text. They even share a wire format: Laya's server exposes the same POST /v1/systemone request and response shape as Jev.

So which should you use? This comparison uses published numbers only, and it says plainly where Jev is ahead. Laya Studio is an independent hosted service for the open-source Laya model. We are not affiliated with TypeSafe AI or with Convai Innovations, Laya's authors.

The one-paragraph verdict

Choose Laya when latency, open weights, European data residency, non-English traffic or per-question pricing matter, and your choice questions have fewer than about 20 options. Choose Jev when you need 20 to 255 options in a single question, very long inputs, or the strongest zero-tuning quality on fine-grained label sets. Because both accept the same request, you can run both in parallel on your own traffic before you commit.

Side by side

Jev figures are third-party published numbers quoted on the Laya model card. Laya's authors note they never measured Jev directly, and sample sizes and prompts differ. Treat the accuracy rows as indicative, not a controlled benchmark.

Where Laya is ahead

Speed

Laya answers every question about a state in one forward pass of an encoder model. On a Tesla T4, in-process, a single question takes 39.5 ms on the English checkpoint and 32.8 ms on the multilingual one. Through Laya Studio's hosted API, a warm request is about 120 ms end to end from Europe, network included. Jev has been measured at 236–276 ms p50 end to end.

These are not like-for-like numbers: network paths and load differ. They do show the same pattern. If a decision sits on the hot path of an AI agent or a user-facing screen, the gap matters. See latency budgets for agents.

Calibration

A probability is only useful if it is honest. After temperature fitting, Laya's expected calibration error is 0.081, against 0.246 published for Jev. In practice, a 0.9 from Laya is closer to "right nine times in ten", which makes confidence thresholds for automation safer. The caveat is that Laya ships over-confident (ECE 0.466) before that fitting. Laya Studio applies the calibration, and self-hosters should too. More in temperature scaling.

Open weights and residency

Laya's weights are Apache-2.0. You can run them on your own hardware, audit them, or fine-tune them. Through Laya Studio, requests are processed on GPUs in Switzerland, the content of a request is never written to disk, and a Swiss-only mode guarantees a request is never answered outside the country. Details are on the Swiss data residency page.

Languages

Laya routes each request by script and language. On the MASSIVE intent benchmark, routing lifts the number of usable languages (more than 3x better than random) from 23 to 45 of 51. Jev has no published multilingual benchmark.

Where Jev is ahead

Many options

This is the biggest difference, and it favours Jev. On Banking77, a 77-way intent task, Laya scores 0.425, while Jev is reported at 0.870 on a 72-label version. Jev supports up to 255 options out of the box. Laya's decision head gives each option only a few tokens once the list gets long, so accuracy drops.

If you have a large taxonomy, you can work around it with Laya by asking a coarse question first and a fine one second. But if a single 100-option question is core to your product, Jev is the stronger choice today.

Fine-grained ratings

Laya's ordinal score questions are its weakest type (SST-5, five sentiment levels: 0.372). On the typed-decisions benchmark, Jev's soft-probability accuracy (0.580) beats Laya's specialist checkpoint (0.471).

Long inputs

Laya reads up to 512 tokens per question on the English model and 1,024 on the multilingual one. Jev accepts much longer states. For full documents, chunk first or use Jev.

What about price?

Both bill the same unit, input tokens, so the comparison is direct.

• Jev bills input tokens: $0.042 per million (list price).
• Laya Studio bills input tokens too: $0.0294 per million, 30% below Jev's list price. Every new account gets 5 free runs. Output is never billed because nothing is generated, and failed requests are free.

A worked example: one million requests of about 120 input tokens each is 120 million tokens, or about $5.04 on Jev's list price and about $3.53 on Laya Studio. One caveat: the two models use different tokenizers, and Laya counts the text once per question it answers, so token counts for the same request are close but not identical. Model your own mix on the pricing page.

Migrating from Jev to Laya

Because the request and response shapes match, a migration is mostly a configuration change:

# Before (Jev)
curl https://api.typesafe.ai/v1/systemone -H "Authorization: Bearer $TYPESAFE_API_KEY" -d @request.json

# After (Laya Studio): same body
curl https://api.laya.studio/v1/systemone -H "Authorization: Bearer $LAYA_API_KEY" -d @request.json

A model field naming a Jev model is ignored, and Laya picks the right checkpoint for the language. Before you switch any traffic:

1. Send a sample of real requests to both and compare answers.
2. Check any question with more than about 20 options. Split it or keep it on Jev.
3. Set confidence thresholds from Laya's probabilities, not Jev's, since the two calibrate differently.

The full checklist is in the Laya vs Jev comparison, and background on Jev is in what is Jev?.

FAQ

Is Laya a drop-in replacement for Jev?

On the protocol, yes: Laya speaks Jev's /v1/systemone request and response shape, so existing clients work after changing the base URL and key. On accuracy, test first, especially on questions with many options, where Jev is stronger.

Is Laya Studio affiliated with TypeSafe?

No. Laya Studio is an independent service built on the open-source Laya model. It is not affiliated with or endorsed by TypeSafe AI or Convai Innovations.

Can I use both?

Yes. Since the requests are identical, you can route each question type to the model that handles it best, for example many-option intent questions to Jev and fast yes/no gates to Laya.

Which is faster?

On published measurements, Laya. A warm Laya Studio request takes about 120 ms end to end from Europe, and Jev has been measured at 236–276 ms p50. Measure from your own region before deciding.