Jev vs Laya: Hosted API or Open-Source Model?

Laya is the open-source model people compare Jev with most. It appeared on GitHub on September 18, 2026, three days after Jev, by Nandakishor M of Convai Innovations, whose Hacker News post “I built non-autoregressive decision models with RL a year ago” reached 1,351 points. A week later its GitHub repository had about 23,700 stars.

Laya does not copy Jev’s weights, which are not public. It is a separate, much smaller model that answers the same kinds of questions in the same format.

At a glance

	Jev	Laya
Made by	TypeSafe AI	Nandakishor M / Convai Innovations
Released	September 15, 2026	September 18, 2026
Access	Hosted API only (TypeSafe, OpenRouter, Vercel, Cloudflare)	Download and run yourself: pip install laya
License	Proprietary	Apache-2.0
Model size	Not published	421M (English, ModernBERT-large); 322M (multilingual, mmBERT-base)
Price	$0.042 per million input tokens, output free	Free; you pay for your own hardware
Question types	Noul, Choice, Score	Noul, Choice, Score, same JSON shape
Input length	About 32k tokens of state plus the longest question	512 tokens (English checkpoint); 1,024, up to 8,192, on the multilingual one
Options per Choice	Up to 255	Limited by a shared token budget for option text; accuracy drops beyond about 20 options
Languages	Best in English	Router sends non-English text to a multilingual checkpoint covering 100+ languages
Fine-tuning	Not offered	Yes, with a notebook that runs on free Kaggle GPUs
Data leaves your machine	Yes	No

Our head-to-head test

On September 25, 2026 we ran both models on the same 200 questions from two public datasets: 100 news articles from the AG News test set (choose one of 4 topics) and 100 customer messages from the Banking77 test set (choose one of 77 intents). Each model got exactly the same instruction and option descriptions, one Choice question per item. Jev was jev-1.13.0 on TypeSafe’s API. Laya was version 0.3.20 with its default router, which picked the English checkpoint for every item, run on an Apple M2 Max laptop. We did not tune or calibrate either model.

Result	Jev	Laya
AG News, 4 options: correct out of 100	86	96
Banking77, 77 options: correct out of 100	78	41
Median time per question, AG News	523 ms (API call from East Asia, includes network)	34 ms (laptop GPU); 169 ms on CPU only
Median time per question, Banking77	552 ms (same)	54 ms (laptop GPU)
Average probability of the chosen option when wrong, AG News	0.87	0.65
Average probability of the chosen option when wrong, Banking77	0.72	0.84

Same questions and option descriptions for both. Jev time includes the network trip from East Asia; Laya ran on an M2 Max GPU.Replay

What the numbers say:

• Few, clearly different options: Laya is at least as good. On AG News, Laya got 12 items right that Jev missed, and Jev got 2 that Laya missed. Laya was also less sure of itself when it was wrong, which is what you want from a probability.
• Many similar options: Jev is far ahead. On Banking77, Jev got 39 items right that Laya missed, and Laya got only 2 that Jev missed. Laya was also confidently wrong: its wrong answers averaged 0.84 probability. Laya’s own README explains why: all option descriptions share a fixed token budget, so with 77 options each label gets only a few tokens.
• Speed depends on where the model runs. Jev’s times include a trip across the Pacific from our test machine; independent US-based tests cited in Laya’s README measured a median of about 240 to 280 ms. Laya runs next to your code, so there is no network at all.

These are small samples on two tasks with one wording, so treat them as a direction rather than a ranking. They match the pattern in Laya’s own README, which reports Laya ahead on AG News and on emotion labels, and Jev far ahead on Banking77. Our full scripts, samples and raw outputs are kept in our test records.

What Laya’s own benchmarks claim

Laya’s README compares it with Jev on several public and synthetic tasks, and is unusually open about the limits:

• Its Jev numbers come from third-party published runs, not its own Jev calls, so prompts and sample sizes differ.
• The headline typed-decisions score (0.766 against Jev’s 0.727) comes from a separately fine-tuned checkpoint. The README says the base checkpoints score below a majority-class baseline on that benchmark, and that all of the capability there comes from fine-tuning.
• Its best calibration figure (0.081 expected calibration error) is after fitting temperatures on your domain; as shipped, the README calls the checkpoints over-confident.
• It lists where Jev leads: more than about 20 options, matching full probability distributions, and raw calibration before any fitting.

Switching between them

Laya ships a server, laya-serve, that speaks the same POST /v1/systemone protocol as TypeSafe’s API. Existing Jev code can call it by changing the base URL, for example with the official Python SDK’s base_url setting. Three things differ when you move code across, according to Laya’s documentation:

1. Options per question share a token budget, so long or numerous option descriptions get cut. Jev accepts up to 255 options.
2. Score levels all need descriptions; a null level is rejected with 422.
3. confidence is calculated differently, so a threshold tuned on Jev does not transfer. Laya suggests gating on its answer_confidence field instead.

Because both accept the same request, running your own labeled examples through each is the fastest way to decide. See Your first Jev call for the request format.

Which should you use?

Your situation	Better fit
A few well-separated options (topic, sentiment, yes/no checks) at high volume	Laya, if you can host it
Dozens of similar labels, such as support intents or product categories	Jev
Long documents, logs or conversations as the state	Jev (32k tokens against Laya’s 512 to 8,192)
Data that must not leave your servers	Laya
Non-English text	Try Laya’s multilingual checkpoint; Jev is strongest in English
You have labeled data from your own domain	Either: fine-tune Laya, or use the data to set thresholds on Jev
No machine learning infrastructure at all	Jev

Laya is one of more than a dozen open-source Jev alternatives; the others, including Kev, SemIf and Bespoke Nimble, are compared on Jev alternatives. For Jev’s own weak spots, see Jev limitations.