Laya Studio FAQ: pricing, accuracy, Jev and Swiss hosting

You send a piece of text and a few questions about it ("Is this urgent?", "Which team should handle it?"), and Laya Studio sends back an answer to each question with a probability that says how sure it is.

No. The live demo on the homepage runs without an account or key: paste some text, ask a question, see the answer. Putting it into your own product takes a developer a few lines of code.

A hosted, autoscaled API for Laya, the open-source (Apache-2.0) System 1 decision model published by Convai Innovations. You get an endpoint, API keys, usage tracking and billing; the model weights are exactly the published ones. Laya Studio is independent and not affiliated with Convai Innovations or TypeSafe.

It reads a state (text, JSON or a conversation) together with typed questions and scores every option of every question in a single forward pass of a bidirectional encoder. It returns a probability distribution per question. It never generates text, so it cannot make up an answer that is not one of your options.

The term borrows Kahneman’s fast/slow distinction: a System 1 model makes quick, intuitive, well-defined decisions in one pass, whereas System 2 (an LLM reasoning step by step) is slower and more flexible. Laya is for the former. See /learn/system-1-vs-system-2-ai.

The Laya package implements the Jev /v1/systemone request and answer schema, and Laya Studio exposes it at https://api.laya.studio/v1/systemone. Existing Jev clients work after changing the base URL and key. Model quality and confidence semantics differ, so re-validate thresholds.

From the model card: AG News 0.950, XNLI English 0.860, typed-decisions 0.766 with the fine-tuned checkpoint. Weak spots: many-label questions (Banking77 0.425) and ordinal scores (SST-5 0.372). Test on your own data before automating.

Laya is trained with proper scoring rules, but the published checkpoints are over-confident before per-task temperature fitting (mean ECE 0.466 → 0.081 after fitting, per the model card). Use confidence to rank and set thresholds from your own labelled sample.

A warm single request takes about 120 ms end to end from Europe. One RTX 4090 handles roughly 1,000 decisions per second.

The English checkpoint reads 512 tokens per question, the multilingual and typed-decisions checkpoints 1,024. Longer states are truncated, so put the decisive content first.

English on ModernBERT-large, 100+ languages on mmBERT-base. Requests are routed automatically by script and language; pass "lang" to override when detection is unsure, e.g. for short messages.

$0.0294 per 1M input tokens — 30% below Jev: Jev lists $0.042 per 1M input tokens, so the prices compare directly. 1 credit = 1 input token, the tokens the model reads (the text is read once per question, as in Jev). 5 free runs on signup, then credit packs or monthly plans from $19. Failed requests are free.

No. Every new workspace gets 5 free runs: its first 5 successful requests are free. After that you add a card to buy credits or a plan.

Requests are micro-batched on GPU workers and load-balanced across pools; when one pool is saturated or unhealthy, traffic fails over to an autoscaling cloud pool automatically, unless your workspace has Swiss-only mode on. The API gateway runs on Cloudflare’s edge.

Not on self-serve plans yet. Live health is at /status. Contact hello@laya.studio for committed capacity.

Our primary inference pool runs on dedicated GPUs located in Switzerland. Every API response tells you where it was processed in the x-laya-region header. Accounts, API keys (stored only as SHA-256 hashes), credit balances and usage metadata live in a Postgres database in the AWS Zurich region (eu-central-2).

The text and questions you send are processed in memory and discarded when the answer is returned. They are never written to a database or log, and never used to train anything. One exception to know about: the optional chat agent (POST /v1/ask, the MCP ask tool, dashboard Chat) first passes your prompt to a third-party language model to plan the questions, in the US by default and in Switzerland under Swiss-only mode.

One switch per workspace (or the x-laya-residency: ch header per request) and requests are only ever answered in Switzerland. If the Swiss pool is unavailable you get an error, never a silent detour abroad. It is an add-on: +15% credits per request, still about 20% below Jev.

For billing and debugging we keep request metadata only — time, status, number of questions, latency — for 30 days. Daily totals are kept for invoicing.

All traffic uses TLS. The API gateway runs on Cloudflare’s network, which forwards requests encrypted and does not store their content; the GPU servers accept requests only from our gateway.

No retention plus Swiss processing makes Laya Studio a fit for sensitive text such as patient messages, clinical intake notes, HR cases or financial correspondence — the kind of data that should not sit in a US AI provider’s logs. Typical uses: Patient message triage, Clinical intake routing, Pharmacovigilance signal screening, HR and legal case sorting, Insurance claim classification. You remain responsible for your legal basis to process personal and health data. Laya Studio holds no formal certification (such as ISO 27001) today and does not sign HIPAA BAAs.

Designed to support compliance with the Swiss Federal Act on Data Protection (nFADP, in force since 1 September 2023) and the EU GDPR. Our Data Processing Agreement (laya.studio/dpa) is part of the Terms and applies automatically when you send personal data; a signed copy is available on request: privacy@laya.studio. You remain responsible for your legal basis to process personal and health data. Laya Studio holds no formal certification (such as ISO 27001) today and does not sign HIPAA BAAs.

You remain responsible for your legal basis to process personal and health data. Laya Studio holds no formal certification (such as ISO 27001) today and does not sign HIPAA BAAs.