Skip to content
TrustList
L
AI Model

Laya

Open-weight Apache 2.0 model that answers typed questions with probabilities in one pass on local hardware

About Laya

Convai Innovations published three open-weight "System 1" decision models, grouped as Laya, on 18 and 19 September 2026 under Apache 2.0. None writes text. The input is a state, such as an email, a ticket or a block of JSON, together with typed questions, and the output for each question is a choice, a score or a probability, produced in one forward pass. That puts it alongside other small decision models here as other small decision models in this directory, including Jev from TypeSafe and CUA-S1-FORMS from Cua.

The three checkpoints

In the English checkpoint, a ModernBERT-large encoder feeds a compact decision head that was trained from scratch; the total is 421 million parameters, with 512 tokens available for each question. The multilingual checkpoint swaps in an mmBERT-base encoder, giving 322 million parameters and 1,024 tokens. A third checkpoint is tuned for four typed-decision workflows. The authors train with a method they name RLCD, reinforcement learning against strictly proper scoring rules, designed so that stating honest probabilities is the only way to earn reward.

Speed, as the authors measured it

On a Tesla T4 the authors measure around 39.5 milliseconds per question for the English checkpoint and 32.8 for the multilingual one, dropping to 7.2 milliseconds per question when ten are batched. No benchmark score is stored for them, and the headline comparison with Jev blends the authors' own measurements with third-party Jev figures taken on different prompts and sample sizes, which the model card itself says.

Where the model card shows it fails

The release earns attention for something most vendors omit. By the card's own account, the English checkpoint breaks down on non-Latin scripts and yet stays sure of itself, scoring zero accuracy with very high confidence on Khmer, which means a confidence gate will not catch the failure. The multilingual checkpoint is shipped uncalibrated and is consistently over-confident, so a buyer should refit its temperature on their own data before trusting the probabilities. For this reason a router selects the checkpoint from the script of the input ahead of the forward pass.

Running it

There is no hosted API and no price: Laya runs on the buyer's own hardware. The company says the underlying research goes back to March 2025; the date used in this entry is the one on which the downloadable checkpoints first appeared.

TrustList summary

Laya is a family of three open-weight decision models from Convai Innovations, published under Apache 2.0 in September 2026. Give it a state (an email, a ticket, or some JSON) and typed questions, and it returns a choice, score or probability for each question in one forward pass. It is built for teams that want to run decision models on their own hardware, since no hosted API is offered.

What Laya does

  • English checkpoint: a ModernBERT-large encoder with a small decision head, 421 million parameters in all.
  • Multilingual checkpoint: an mmBERT-base encoder of 322 million parameters, with a 1,024-token budget per question.
  • Workflow checkpoint: tuned on four typed-decision workflows.
  • Script router: chooses the checkpoint from the script before the forward pass.
  • Timing: the authors report 32.8 ms per question on a Tesla T4 for the multilingual checkpoint, measured by themselves.

Who it works with

Teams that run decision models on their own hardware and can test the checkpoints on their own data.

Key facts

  • Licence: Apache 2.0
  • Release: downloadable checkpoints appeared on 18 and 19 September 2026
  • Research: the company says the underlying work dates from March 2025
  • Caveat: the model card says the English checkpoint fails on non-Latin scripts while staying confident, and that the multilingual checkpoint is over-confident until its temperature is refitted on the buyer's own data

Summary written by TrustList from the business's own published information, Nov 2026.

Benchmarks & AI stats

BenchmarkOfficialCommunity avg
No benchmark scores yet. Be the first to add one.

“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).

Sign in to add a benchmark score for this model.

Request a demo or quote from Laya

Protected by reCAPTCHA - Google Privacy Policy and Terms apply.

By sending, you agree we may share your request and contact details with the provider once you confirm your email.

Reviews

Write the first review of Laya

Used it? Your experience helps other buyers decide.

Write a review

Questions & answers

No questions yet. Be the first to ask about Laya.

Laya alternatives

Similar listings buyers compare against this one.

Microsoft AI's text-to-speech model for keeping one voice across 23 languages, with voice cloning from short audio

Low-latency version of ElevenLabs' Eleven v4 speech model, aimed at phone and chat voice agents in 90+ languages

Open-weight Qwen-based model from Bad Theory Labs for terminal and software engineering agents, under the Qwen Community License

Open-source 2B decision model from AWS's Strands Agents project that picks options and returns scores with a reliability figure

Microsoft AI's real-time speech-to-text model for voice agents, transcribing 60 languages with continuous language detection

Microsoft AI's high-volume, low-latency variant of MAI-Voice-2.1 for speech in 23 languages, aimed at voice agents

Lower-cost sibling of Cohere's Embed 5 Pro, for high-volume multimodal embedding and indexing across 100+ languages

Ai2's open-weight 8B model that writes cited scientific reports from a research question and retrieved literature