Skip to content
Jev Pricing

Independent comparison · not affiliated with TypeSafe AI or Convai Innovations

Jev vs Laya: API cost vs self-hosted Laya

Laya is an open-weight decision model you run yourself; Jev itself cannot be run locally. For 400K decisions a day, Jev costs $126.00 a month and one always-on Google Cloud L4 costs $509.04. On price alone, one GPU breaks even at about 1.6M decisions a day, if it can keep up. Laya can still win when data must stay local, when you need answers in milliseconds, or when the hardware is already yours.

  • Cost: Jev bills per token; a GPU bills every hour. Low volume favours Jev, very high volume a GPU.
  • Latency, not head to head: Laya took 7.8–17.5 ms for a short local request on our Mac; Jev took 375 ms for a ~1,000-token request over the network.
  • Context: Laya reads 512 or 1,024 tokens (up to 8,192 configured); Jev 64K total direct, about 32K via gateways.
  • Accuracy, one third-party test: at 128 candidate answers, Jev 60% and Laya 39%. Not a general ranking.
Your numbers

Which is cheaper for your volume?

Change the volume and where Laya would run. Prices are on-demand list prices, checked 2026-09-26.
Jev API, per month
$126.00
Laya, always on
$509.04
minimum: 1 instance
Decisions per day
400K
Price break-even, one instance
~1.6M decisions/day
One instance handles about
enter throughput below

Each decision is about 560 tokens (state + question), within Laya multilingual 322M's 1,024-token limit. Jev allows 32K via gateways, 64K direct.

An instance is billed every hour of the month, busy or not: Google Cloud L4 (g2-standard-4), $0.707/h on-demand. Above the break-even, cost still steps up by one instance at a time. 30-day month, one batched Jev call per item, including the ~260 tokens every Jev call bills; Laya re-reads the state for each question.

Choose Jev when

  • volume is below the break-even, or it swings
  • requests carry long text (over 1K tokens)
  • you would rather not run servers
Every way to call Jev, with working calls →

Choose Laya when

  • data must not leave your machines, or must work offline
  • a few milliseconds per answer matters
  • you already own the hardware, or your own GPU benchmark beats the bill
Check your request’s size first →
Measured

Laya-MLX on an Apple M4 Pro, 20 GPU cores, 64 GB

Three independent runs per model, 200 samples per test, plus five minutes of sustained load. Latency includes tokenization and formatting, not loading.
P50, end to endLaya multilingual 322MLaya 421M (English)
One short question7.8 ms17.5 ms
~420-token text, one question23 ms57.1 ms
50 questions in one call183 ms519 ms
Sustained, questions/s107.442.7
Start to first answer0.62 s0.36 s
Peak memory, one question688 MiB944 MiB

The port’s author reports 7.39 ms and 13.42 ms for the same short question on an M3 Max, which has twice our GPU cores. One thing the table hides: Laya reads the text again for every question, so 50 questions cost 50 times the text. Jev bills the text once per call, which is why batching questions cuts its bill and not Laya’s work.

Before you switch

What changes besides the bill

Accuracy

In one independent test (classification with 64 to 128 candidate answers), Jev answered 60% correctly at 128 candidates and Laya 39%. At 64, hard distractors took Jev from 96.8% to 86.3% and Laya from 90.5% to 56.0%. Reordering the options changed 49.4% of Laya’s answers and 14.6% of Jev’s, though Jev also varied on 4.3% with the order held fixed. A high-cardinality test, not a general ranking; we did not measure accuracy ourselves.

Context

The English model reads 512 tokens, the multilingual 1,024 (its author allows up to 8,192, with accuracy that varies beyond about 4,000). Text past the limit is cut. Jev takes 32K through gateways and 64K direct: estimate your request.

Drop-in, almost

Laya’s laya-serve speaks the same POST /v1/systemone protocol as Jev, so TypeSafe’s SDK can point at it. Choice options share a 192–256-token budget rather than Jev's 255 options, and confidence is computed differently, so a threshold tuned on Jev does not carry over.

Known issues

  • The multilingual checkpoint rarely picks the first-listed level of a Score question, in any language. Source
  • The multilingual checkpoint is weaker on English than the English one (0.619 vs 0.684 macro across English suites). Source
FAQ

Jev vs Laya questions

Is Laya cheaper than Jev?

Only at high volume, or on hardware you already own. At 400K decisions a day (100K items, 4 questions each, 500-token state) Jev costs $126.00 a month; one always-on Google Cloud L4 costs $509.04. On price alone one GPU breaks even at about 1.6M decisions a day, if it can keep up.

Is Laya as accurate as Jev?

Not in the one independent head-to-head we found, which tested classification with 64 to 128 candidate answers: Jev answered 60% correctly at 128 candidates and Laya 39%. That is a high-cardinality test, not a general ranking; test on your own labelled data before switching.

Can Laya replace Jev without code changes?

Partly. Laya's laya-serve speaks the same POST /v1/systemone protocol as TypeSafe's API, so an SDK pointed at it works, but options share a 192–256-token budget, confidence is computed differently, and the context is 512 or 1,024 tokens instead of Jev's 32K–64K.

How fast is Laya on a Mac?

On an Apple M4 Pro, 20 GPU cores, 64 GB, one short question took 7.8 ms with the multilingual model and 17.5 ms with the English one; sustained, it answered 107.4 and 42.7 questions a second with 3 questions per call. Measured with Laya-MLX on 2026-09-26.

Can I run Jev itself locally?

No. Jev is TypeSafe's closed model, served only through its API and gateways. Laya is a different, open-weight model with a compatible request format.

Get an email when Jev’s price or limits change

The break-even above moves with Jev's price. We check it every day and will tell you when it changes.

Optional

This site is independent of TypeSafe and cannot give you an account.

How we handle your email
  • What: your email, the boxes you ticked, the page and the date. Your IP address is not stored.
  • Why: only to send the notices you asked for, on your consent. Never shared or sold.
  • Where: a Cloudflare D1 database. Cloudflare Turnstile checks the form for bots and processes browser signals to do it (its privacy policy).
  • How long: until you remove it. The first email asks you to confirm; addresses not confirmed within 30 days of it are deleted.
  • Your rights: remove your address with the link above at any time, or write to [email protected] to see, correct or delete what we hold.

Laya facts from its repository and model cards (Apache 2.0 (code and weights)), checked 2026-09-26. Mac measurements: jev-fanout-bench round 3, 2026-09-26, with Laya-MLX; no Jev call was made for them. Jev at $0.042 per 1M input, checked 2026-09-25. GPU prices are on-demand list prices, checked 2026-09-26. Laya’s training data is not fully published; we found no evidence it was trained on Jev’s outputs. Independent of TypeSafe AI and Convai Innovations.