Jev vs Laya
Laya is an open-source encoder family you host and fine-tune. Jev is a hosted decision model you call. Weights vs API: what changes, what it costs, and when each is the right call. Independent hosted gateway; not the official TypeSafe site.
What Laya is
Laya, from Convai Innovations, is Apache-2.0. It puts choice, score and yes/no heads on ModernBERT and mmBERT encoders, so it runs on a CPU, a consumer NVIDIA GPU or Apple Silicon; a multilingual checkpoint covers more than English. You pull the weights, wire the server, and own your fine-tunes.
Your weights, not a hosted call
A Laya base checkpoint is a model you fine-tune on your own labels — it is not a copy of Jev's weights, and a zero-shot base will not stand in for hosted Jev on your task until you train it. You pay in machines and labeling time, not in per-request credits; in exchange, your data never leaves your machines and there is no vendor account to depend on.
The cost picture
Hosting is not free: an always-on GPU instance, capacity headroom, and the labels you collect before the model learns your task. The API path charges per use — 1 credit per 1,000 input tokens, $10 for 100,000 credits that never expire — with nothing to operate. A working rule from teams comparing the two: below a few thousand decisions a day the API wins decisively; at sustained high volume, with labels in hand and data residency constraints, the fine-tuned road becomes rational.
When to pick which
Pick Laya when the labels are yours, the data cannot leave your machines, or the task is stable enough that a trained model amortizes well. Pick Jev through Optio when you want a hosted decision today — keys and prepaid credits in minutes, no GPU care, and calibrated probabilities that work on a new task shape the day it appears, without a retraining cycle.
Strengths and honest limits
Jev's honest limits: text only, eight questions per request, and no fine-tuning — your only lever is the wording of instructions and criteria. Its strength is the flip side of the same coin: no labeled set to collect, no training run to outlive your product, and criteria that change in one afternoon when the decision changes. If the task shape keeps moving, that is usually the winning side of the trade.
Quick comparison
| Question | Jev | Laya |
|---|---|---|
| Hosting | Managed — Optio runs it, you call it | You host (CPU / GPU / Apple Silicon) |
| License | Proprietary via API access | Apache-2.0 open weights |
| Setup | Create a key, one REST call | Pull weights, fine-tune on your labels, run servers |
| Cost shape | Per use: 1 credit ≈ 1,000 input tokens; $10 = 100k credits | Compute + labeling up front; near-zero marginal cost after |
FAQ
- Is Laya the same model as Jev open-sourced?
- No. Laya is a different model family — encoder heads on ModernBERT/mmBERT from Convai Innovations, Apache-2.0. It targets the same question types (choice, score, yes/no) but is an independent project with independent weights and independent behavior on your data.
- Which is better for languages other than English?
- Laya ships a multilingual checkpoint, which is its strength; Jev is a text model whose instructions and criteria you write in your language. If your decisions run heavily in one non-English language and you can fine-tune, try Laya's multilingual checkpoint on your labels; for mixed or changing language sets, start with the hosted English-first path and test.
- Which is cheaper — credits or self-hosting?
- At low volume, credits: $10 covers 100,000 short decisions and you operate nothing. At high sustained volume with access to labels and GPUs already paid for, a fine-tuned Laya can reach a lower marginal cost — the crossover usually sits somewhere above tens of thousands of decisions per day, and only after your labeling investment is done.
- Is this the official Jev site?
- No. Optio is an independently operated gateway for calling Jev and is not affiliated with TypeSafe or with any alternative model mentioned on this page.