Compare · Last verified 2026-10-22

Jev vs Laya: pricing, limits, and which to use

Verdict

Jev (TypeSafe AI) is a hosted API at $0.042 per 1M input tokens with free output and a 64k-token context (262k opt-in), checked 2026-10-22 on docs.typesafe.ai/models. Laya is a free Apache-2.0 open-weight model you host yourself: the English checkpoint reads 512 tokens and the multilingual and typed-decisions checkpoints 1,024. Laya’s model card reports it is faster than Jev and more accurate on several small-label benchmarks, but far behind on a 77-label task. Those are the Laya authors’ numbers (Jev figures come from third-party benchmarks they cite), and we have not reproduced them.

Pick Jev for: Long state (64k tokens, 262k opt-in), questions with dozens of options, and teams that want a managed API with nothing to host.
Pick Laya for: Self-hosted or offline use, short inputs, low latency on your own CPU or GPU, and no per-token bill (you pay for compute).

Side-by-side

	Jev (TypeSafe AI)	Laya (open source)
Vendor	TypeSafe AI	Convai Innovations
Category	Decision model	Open-weight decision model
Price	$0.042 per 1M input tokens ($42 per billion) for both Jev 1.14 and 1.13; output tokens free. Long-context mode costs the same per token. [Verified 2026-10-16 on docs.typesafe.ai/models (no change)]	Free open weights (Apache-2.0); you pay your own compute. [Verified 2026-10-15 on Hugging Face model card]
Limits	64k tokens per request / 32k per question by default; 262k for both with the opt-in long-context header (slower and less accurate per TypeSafe). [Verified 2026-10-16 on docs.typesafe.ai/models (no change)] English is the primary training language and where accuracy is best; other languages (including CJK) are handled less well, per TypeSafe. [Verified 2026-10-16 on docs.typesafe.ai/models]	512-token context on the English laya checkpoint (about 320 tokens left for state at default settings). [Verified 2026-10-22 on Hugging Face model card] Two more checkpoints in the same repo: laya-multilingual (mmBERT-base, 322M, 1,024 context, encoder up to 8,192) and laya-typed-decisions (421M, 1,024 context). Default option budget gives 77-option questions only 3–4 tokens per label, so accuracy falls on many-option questions unless head_max_len is raised. [Verified 2026-10-22 on Hugging Face model card]
Status	Correction: jev-latest resolves to jev-1.13.0, the most recent stable release. jev-1.14.0 (available from October 14, 2026) is preview only, via the jev-preview alias. [Verified 2026-10-16 on docs.typesafe.ai/models (no change)] Not trained on customer requests; ZDR for enterprise. No SLA stated on the models page. [Verified 2026-10-15 on docs.typesafe.ai/models] Now powers OpenRouter Armor (beta, launched 2026-10-15), per OpenRouter’s launch post. Armor checks agent tool calls against plain-English rules and is free during the beta. See the OpenRouter Armor page linked below. [Verified 2026-10-16 on OpenRouter’s official X posts (2026-10-15)] Also on OpenRouter: typesafe/jev-1.14 and typesafe/jev-1.13, billed to your OpenRouter account, via POST /api/alpha/decisions or /api/v1/systemone. On OpenRouter, ~typesafe/jev-latest tracks the newest release (1.14); TypeSafe’s own jev-latest points to 1.13.0. [Verified 2026-10-16 on openrouter.ai/docs/guides/community/jev] Correction: an earlier version of this page showed $0.42/1M taken from jevtypesafeai.com, a domain we could not tie to TypeSafe. TypeSafe’s own docs show $0.042. [Correction logged 2026-10-15]	Hugging Face model and GitHub repo were both created on 2026-09-18 (UTC), which matches the third-party release date. [Verified 2026-10-16 on https://huggingface.co/api/models/convaiinnovations/laya and https://api.github.com/repos/NandhaKishorM/laya/releases] Latest GitHub release v0.4.13, published 2026-10-21 (runtime fixes; the model card says checkpoints are unchanged). [Verified 2026-10-22 on https://api.github.com/repos/NandhaKishorM/laya/releases/latest] Our smoke test (script: /workspace/research/bench/laya_bench.py): local CPU on an 8-vCPU box, median 181 ms and p95 245 ms per call, 10/10 correct on easy synthetic items, stable across 5 repeats and reversed option order. Small smoke test only, not a quality benchmark. Confidences are uncalibrated (the loader warns the checkpoint ships invalid temperatures). [Verified 2026-10-15 — AgentAPIList local run]
Fit notes	Pin a versioned ID (e.g. jev-1.13.0) rather than jev-latest or jev-preview so thresholds don’t drift with version changes. Text-oriented; check image support for your version in TypeSafe docs before relying on it.	Tiny and fast enough for CPU routing and gating; recalibrate before using confidence thresholds. Short 512-token context; truncate or summarize long state. The model card says base checkpoints are near chance on typed-decisions zero-shot; its 0.766 score comes from the checkpoint fine-tuned on that benchmark. Plan to fine-tune for your task.

Benchmarks (vendor-reported)

Reported by the Laya authors on the Hugging Face model card, checked 2026-10-22. Laya figures are from Laya’s Router; Jev 1.13.0 figures are third-party published numbers the authors cite, not measured by them. Sample sizes and prompts differ. Not reproduced by AgentAPIList. Source: Laya model card: Laya vs TypeSafe Jev.

Metric	Jev	Laya
typed-decisions accuracy (2,000 decisions)	0.727	0.766 (fine-tuned laya-typed-decisions; base laya scores 0.362)
AG News accuracy (4 labels)	0.910	0.950
DAIR Emotion accuracy (6 labels)	0.480	0.595
Banking77 accuracy (Jev 72 labels, Laya 77)	0.870	0.425
Calibration error, ECE (lower is better)	0.246	0.081 after temperature fitting
Median latency, 1 question	236–276 ms	32.8 ms (single T4 GPU)
Weights	Closed, hosted API	Open, Apache-2.0

FAQ

Is Laya better than Jev?

It depends on the task. On Laya’s own model card (checked 2026-10-22), Laya scores higher on AG News (0.950 vs 0.910) and DAIR Emotion (0.595 vs 0.480), and its fine-tuned typed-decisions checkpoint beats Jev 1.13.0 (0.766 vs 0.727). Jev leads on 77-label Banking77 (0.870 vs 0.425), and the authors say base Laya checkpoints are near chance on typed-decisions zero-shot. These are vendor-reported numbers we have not reproduced.

Is Laya free?

The weights are free under Apache-2.0; you pay for your own CPU or GPU. Jev costs $0.042 per 1M input tokens with free output (TypeSafe docs, checked 2026-10-22).

Can Laya replace the Jev API without code changes?

Mostly. Per its model card, laya-serve exposes POST /v1/systemone with the same request and response shape as Jev, so TypeSafe clients can switch by changing the base URL. It has no authentication unless LAYA_API_KEY is set.

Which handles longer inputs?

Jev: 64k tokens per request by default and 262k with the long-context header. Laya: 512 tokens on the English checkpoint and 1,024 on the multilingual and typed-decisions checkpoints (the multilingual encoder can be raised to 8,192).

Methodology

Every fact in the table is copied from the vendor's official page (or OpenRouter where noted) and labeled with the date it was last verified. Items we could not confirm are marked unverified. Prices are list prices per 1M input tokens before regional or volume adjustments. See our methodology.