Laya — System 1 on Base

The engine

Reflex decisions, not a chatbot.

Drawn from the September 2026 research note by Nandakishor Mukkunnoth, ConvAI Innovations. Read the source.

Most pipelines still call a generative model for a label. Laya answers typed questions in one forward pass and never writes a sentence.

2025

The first papers

March 2025 brought arXiv:2503.23303, PPO over sequence representations, turn-by-turn conversion probabilities, the Hugging Face weights sales-conversion-model-reinf-learning, and the open saas-sales-conversations set. September 2025 formalized schema decisions in arXiv:2510.01237. Reinforcement learning stayed the guide, not a plain embedding model.

2026

Then Jev arrived

TypeSafe AI, founded by Diogo Almeida, shipped Jev as a non-autoregressive decision product: RLCD, horizontal confidence distributions, about $0.042 per million input tokens, and responses near 150 ms. No papers, no open weights, no open training set. Laya is the open reply: bidirectional encoders, 32.8 ms on one GPU, 7.2 ms per question when batched, 100+ languages, no API meter.

S1

System 1 versus System 2

Routing a ticket, flagging phishing, scoring urgency from 0 to 3, or spotting a jailbreak does not need an 8B or 70B model to stream tokens for half a second to two seconds. A quoted confidence from an LLM is just more tokens. Laya’s probabilities are the output, so they can be calibrated.

Primitives

Three questions. One pass.

Any state works: raw text, an email, a ticket, or a JSON document. The output is only numbers, so a malformed schema cannot appear.

choice

Pick one key from a dictionary of criteria. You get the winner, the full distribution, and a calibrated confidence.

score

Place the state on an ordinal rubric. You get the expected level, the mass on each rank, and a confidence.

noul

A yes-or-no question. P(true) sits between 0 and 1, and P(false) is one minus that value by construction.

Checkpoints

Three models, one hub.

No single encoder is best at every language and task. All three live under convaiinnovations/laya. The SDK pulls only the subfolder you name, so you skip the full 2.5 GB bundle.

Checkpoint	Encoder	Params	Context	Where it leads
laya	ModernBERT-large	421M	512	English classification, guardrails, email triage
laya-multilingual	mmBERT-base, 256k vocab	322M	1024, up to 8k	100+ languages, about 2.2× faster, cross-lingual NLI
laya-typed-decisions	ModernBERT-large	421M	1024	Agents, support, invoices, security alerts, 0.766 accuracy

agent_en = laya.load("convaiinnovations/laya")          # English, ~808 MB
agent_ml = laya.load("convaiinnovations/laya", subfolder="multilingual")  # ~647 MB

Routing

Confidence will not warn you.

On MASSIVE, 20 options, chance is 0.050. ModernBERT-large’s 50k English vocabulary shreds other scripts, and the model stays sure of itself anyway. Across 51 languages its mean confidence never fell below 0.885, whether accuracy was 82% or zero. Pick the checkpoint before the forward pass.

Script	Accuracy	Mean confidence
Khmer	0.000	0.952
Armenian	0.050	0.885
Hebrew	0.060	0.964
Bengali	0.080	0.945
Hindi	0.100	0.941

0.09 ms

English

Unicode script scan plus Latin stopwords. Overhead stays under 2% of a 33 ms pass.

0.54 ms

Indic

Devanagari and the other 21 alphabets the router knows, including CJK, Cyrillic, Arabic, Hebrew, Tamil, and Thai.

0.73 ms

Nested JSON

A 200-row document. Router(preload=True) keeps the checkpoints resident and skips a 7–10 second cold swap.

Head to head

Routed Laya against Jev 1.13.0.

Laya figures are measured. Jev figures come from TypeSafe and from independent write-ups by AbdelStark and nibzard.

Measure	Jev 1.13.0	Laya, routed	Delta
typed-decisions, 2,000	0.727	0.766	+3.9%, above a 0.735 teacher
AG News, 4 labels	0.910	0.950	+4.0% accuracy
DAIR Emotion, 6 labels	0.480, Brier 0.846	0.595	+11.5%; Jev put 16% at zero probability
Calibration error (ECE)	0.246	0.081	About 3× tighter
Latency p50, 1 question	236–276 ms	32.8 ms	7.8× faster
Latency p50, 10 questions	~1,500 ms serial	72.3 ms, 7.2 ms each	About 20× on a batch
Languages above 3× chance	Unpublished	45 of 51	Coverage, not a single script
Cost per 1M tokens	$0.042 API	$0 self-hosted	Apache 2.0 weights
Weights and code	Closed API	Open safetensors	Runs on-prem and air-gapped

0.993

Enron spam

Accuracy and F1 both 0.993. ECE 0.013.

0.980

Phishing

Accuracy 0.980, F1 0.979, ECE 0.012.

0.931

ToxicChat guardrail

Held-out accuracy 0.755–0.762. At 50% coverage it reaches 0.931.

0.657

RAG relevance

Passage filtering in a single forward pass.

0.522

10-way queues

Support ticket routing across ten destinations.

Ceilings

Where the score stops climbing.

01

More than 20 choices

Banking77 has 77 labels. Laya scored 0.425 there; Jev scored 0.870. Options share a 192–256 token head budget, so each candidate gets only a few tokens. Keep schemas under 20 options, or split them coarse-to-fine.

02

Zero-shot is not the 0.766

Untuned base weights land near 0.35 on typed-decisions, close to chance. The 0.766 number is after fine-tuning on that benchmark’s train split. Treat the release as a fast base to specialize.

03

One temperature per question

Shipped weights use raw temperature logits. A single scalar per question type, fit on your own distribution, moved expected calibration error from 0.466 to 0.081.

Quickstart

Installed in one line.

pip install "laya>=0.3.3"

from laya import Router
router = Router(preload=True)

state = {"subject": "API down since 6 AM", "body": "We lost transactions. Refund the SLA."}
questions = {
    "queue": {"type": "choice", "instructions": "Which queue owns this?",
              "criteria": {"infrastructure": "outages", "billing": "refunds"}},
    "urgency": {"type": "score", "instructions": "How urgent?",
                "criteria": ["low", "medium", "high", "critical"]},
    "churn": {"type": "noul", "instructions": "Is the customer about to leave?"},
}
res = router.predict(state, questions)

English text routes to ModernBERT-large. Hindi and other non-Latin scripts route to mmBERT-base. Pass model="typed-decisions" when you already know the domain.

HF

One billion samples.

Total supply 1,000,000,000 $LAYA

Buy / sell tax 0% / 0%
Chain Base
Liquidity Burned at launch
Presale None

How to buy

Four steps into the pool.

01
Open a wallet

Use a wallet that speaks Base. MetaMask, Coinbase Wallet, or Rainbow all qualify.
02
Fund Base ETH

Bridge or withdraw a little ETH onto Base. Gas here is a whisper, not a toll.
03
Swap for $LAYA

Open Uniswap, select Base, and trade ETH for Laya once the contract is live.
Buy
04
Watch the orbit

Pin the chart. The spiral is easier to read when you can see every sample.
Chart

Join us

The lab door is glass.

Charts, chat, and the buy button. Pull up a stool. The orbit has room.

Buy on Uniswap

Dexscreener

X

Blockscout

Laya Research