Laya model explained: checkpoints, API and Laya vs Jev

The Ollama model page names Convai Innovations as publisher; Hugging Face metadata identifies author convaiinnovations and repository creation on 18 September 2026. This establishes Hugging Face availability, rather than a separately verified announcement date. The model-card metadata states license: apache-2.0.

The model card documents up to 8,192 tokens for the multilingual checkpoint with max_len=8192. It reports accuracy of 0.766 for the fine-tune versus 0.362 for the English base on 2,000 decisions across four workflows. These are author-reported results, not independently verified; Laya typed-decisions is the checkpoint fine-tuned on that benchmark’s training split.

Live counters on 27 October 2026: 57,144 downloads and 5,832 likes for the Hugging Face base repository; Ollama displays 16.4K downloads. These platform counters measure different things and are not quality metrics.

The model card describes choice, score and noul for routing, moderation, guardrails and email triage. Laya scores options in one encoder pass. You can run it locally and fine-tune it for your tasks. It does not generate free-form answers or explanations. The English base defaults to only 512 tokens of context. The card cautions that base checkpoints are near chance on typed-decisions zero-shot, so evaluate your specific workflow with labelled examples.

Source: Benchmark Heaven · JevBench and v1.7.11 release JSON. Displayed method: v1.7 with the 10-second rule; the rows were rescored in v1.7.0. Capability averages Intelligence and Calibration; it is not accuracy in percent. Laya was measured locally on CPU, Jev through TypeSafe’s hosted API. These are different runs and access routes, not same-workload production or speed comparisons. Checked 27 October 2026.

Source: Benchmark Heaven · Jev vs Laya, data release v1.7.21 (published 29 October 2026), method v1.7 with the 10-second rule, rescored from the recorded decision times. Capability is the mean of Intelligence and Calibration. The Capability values match the table above. Laya was measured offline on an evaluator-owned CPU container, Jev through TypeSafe’s hosted API; the model card also cautions that base checkpoints are near chance on typed-decisions zero-shot. These are different runs, not same-workload, speed or cost comparisons, and not a ranking. Checked 29 October 2026.

The upstream documentation shows pip install laya and laya.load("convaiinnovations/laya"). The example below combines those documented building blocks; it has not been executed here. First loading downloads weights. For a self-hosted HTTP server, the GitHub README documents pip install "laya[serve]" and LAYA_DEVICE=cuda LAYA_PRELOAD=1 laya-serve. Review network access and authentication before operating it.

pip install laya
python - <<'PY'
import laya
agent = laya.load("convaiinnovations/laya")
print(agent.predict("Hello World", {
    "greeting": {"type": "noul", "instructions": "Does the text contain a greeting?"}
}))
PY

Ollama lists laya, including laya:421m-english-bf16. Its page requires Ollama 0.40.0 or later and shows this local API call. It has not been tested here.

curl http://localhost:11434/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model":"laya","state":"Hello World","questions":{"says_hello":{"type":"noul","instructions":"Does the state text contain a greeting?","criteria":{"true":"The state text contains a greeting.","false":"The state text does not contain a greeting."}}}}'

Hosted options, checked 27 October 2026: an independent community API lists USD 0.0004 per 1,000 input tokens (USD 0.40 per million), with a one-unit minimum per call and free output. It explicitly disclaims affiliation with Convai Innovations, NandhaKishorM or TypeSafe. Vercel AI Gateway names Laya as a supported decision model. This confirms published offerings; we made no paid calls and have not tested their availability or answers here.

If you prefer to avoid operating a model server and need an EU tier with a public DPA, you can evaluate Decision Models’ existing models. They use a Jev-compatible request shape, their own model IDs and their own limits. Check question counts, input limits, contract and answers on your data. Hosted access alone does not establish better decision quality.