What it is
Laya is an open-weight decision model from Convai Innovations, first published on September 18, 2026 under Apache 2.0. Like Jev, it takes a state and typed questions (choice, score and yes/no "noul") and returns probabilities in a single forward pass instead of generating text. Unlike Jev, it is a small encoder model that runs on a laptop CPU or a modest GPU.
There are three checkpoints. The English one is built on ModernBERT-large (421 million parameters, 512-token context); laya-multilingual uses mmBERT-base (322 million parameters, 1,024 tokens by default and up to 8,192) and covers more than 100 languages; laya-typed-decisions is the English model fine-tuned on a four-workflow benchmark. A Router in the Python package detects the script and language and picks the right checkpoint for each request.
The package (version 0.3.20 at the time of writing) also includes laya-serve, a server that accepts the same POST /v1/systemone requests as Jev, so existing Jev clients can switch by changing their base URL. By default it listens on all network interfaces with no authentication unless LAYA_API_KEY is set.
Our review
Laya's own model card is unusually candid about the main problem. The base checkpoints are near chance on its typed-decisions benchmark, scoring 0.362 against a 0.318 random baseline and 0.461 for always picking the most common answer. The headline 0.766, which beats the 0.727 the README quotes for Jev 1.13, belongs to the checkpoint fine-tuned on that benchmark's own training split.
Independent testing agrees. On the community Jev Decision Index 0.4.1, Laya ranks 216th of 228 entries with a score of 9.21, where 0 is random guessing, against 60.24 for Jev 1.14. In the index's calibration sample it was right 37.7% of the time. The card adds that the English checkpoint scored zero on Khmer while reporting 0.952 confidence, which is why the language router matters more than it first appears.
Where it wins is speed and footprint. Convai reports 32.8 ms for one question on a T4 GPU and about 6.8 ms per question in batches of 50, or 193 to 464 ms per request on a CPU. The Decision Index measured a 5.8 ms median with Laya's fused-kernel fast path and 19.8 ms without it; two entrants are now faster still. A free Kaggle notebook covers fine-tuning and temperature fitting end to end.
The known failure modes are specific. Choice questions with more than about 20 options share a fixed token budget, so on Banking77 the README reports Laya at 0.425 on all 77 labels against 0.870 for Jev on 72, not a like-for-like test. Yes/no questions can follow their option labels rather than the input, and the act_probability field carries no usable signal (AUROC 0.30), so the card says to gate on confidence instead. Few vendors print this much about their own weaknesses, which makes the rest of the card easier to trust than its Jev comparison, where every Jev figure is third-party and was never re-measured by Convai.
Where it earns its place
An ML engineer at a fintech with a few thousand labelled support tickets
Route and tag tickets in under 50 ms without sending them to a third-party API
Fine-tune on the labelled tickets with the Kaggle notebook, refit temperatures, then serve with laya-serve on one T4-class GPU. Base-checkpoint accuracy is poor, so the fine-tune is not optional.
A platform team running an LLM agent pipeline
Add cheap guardrail and escalation checks before tool calls
Several yes/no and choice questions go in one forward pass, about 7 ms each in batches of 50 on a T4 by Convai's figures. Self-hosting costs only the GPU, against $0.042 per 1M input tokens on Jev's API; gate on confidence, not act_probability.
A customer operations lead with support queues in Hindi, Spanish and English
Triage multilingual email without paying per request
The Router sends non-English text to laya-multilingual automatically, and 45 of 51 languages scored above three times random in Convai's tests. Without in-house labels to fine-tune on, accuracy will not justify leaving a hosted decision API.
Before you commit
- Setup
- pip install laya (Python 3.10 or newer); the Router downloads a checkpoint on first use, about 808 MB for English and 647 MB for multilingual. The card says to set USE_TF=0 if loading hangs.
- Integrations
- Optional extras add an HTTP server, an MCP server, LangChain and LangGraph support and ONNX Runtime; laya-serve accepts Jev's POST /v1/systemone requests.
- Data and privacy
- Apache 2.0 weights on your own hardware, so no data leaves it. laya-serve binds 0.0.0.0 without authentication unless LAYA_API_KEY is set, so set a key or firewall the port.
- Cost at scale
- No licence or per-request fees. Convai reports 103 to 332 questions a second batched on one T4, and the fine-tuning notebook runs on Kaggle's free two-T4 tier.
- Limits
- 512 tokens on the English checkpoint (about 320 for the state) and 1,024 on multilingual, raisable to 8,192 with uneven accuracy beyond about 4,000 tokens. Ordinal score questions are the weakest type (0.372 on SST-5), and switching checkpoints with max_loaded=1 costs a 7 to 10 second reload.
What's good
- Apache 2.0 weights; runs on CPU or a small GPU
- About 33 ms per question on a T4, by the vendor's measurement
- Router covers 100+ languages across three checkpoints
- Jev-compatible server and free fine-tuning notebook
What's not
- Base checkpoints near chance on hard decisions, by its own card
- Ranks 216th of 228 on Decision Index 0.4.1
- Ships over-confident; needs temperature refitting
- Server has no authentication unless you set a key
Best for
Teams with their own labelled decision data who want a tiny, self-hosted model to fine-tune for fast routing or guardrails.
Alternatives
Laya: common questions
- Is Laya worth it?
- Very fast and genuinely open, but close to guessing on hard decisions out of the box: a base to fine-tune, not a Jev replacement. We rate it 3.4 out of 5.
- How much does Laya cost?
- Free weights and Python package under Apache 2.0. No hosted paid API; you run it yourself.
- What are the downsides of Laya?
- Base checkpoints near chance on hard decisions, by its own card; ranks 216th of 228 on Decision Index 0.4.1; ships over-confident; needs temperature refitting; server has no authentication unless you set a key.
- Who is Laya best for?
- Teams with their own labelled decision data who want a tiny, self-hosted model to fine-tune for fast routing or guardrails.
- How much does Laya cost for a team or in production?
- No licence or per-request fees. Convai reports 103 to 332 questions a second batched on one T4, and the fine-tuning notebook runs on Kaggle's free two-T4 tier.
- Is Laya safe for company data?
- Apache 2.0 weights on your own hardware, so no data leaves it. laya-serve binds 0.0.0.0 without authentication unless LAYA_API_KEY is set, so set a key or firewall the port.
- What are the best Laya alternatives?
- The closest alternatives we have reviewed are GLiNER2.5-Decide, Jev, Perplexity Decider (pplx-decider), d3 (vLLM Semantic Router) and Clef.
Answers checked Oct 23, 2026.