Convai Innovations · Open-weight decision model · Last verified 2026-10-15

Laya (open source)

Verified

Open-weight 421M encoder decision model (ModernBERT-large), Apache-2.0. laya-serve exposes a System One-style /v1/systemone endpoint; also runs in llama.cpp.

What it does

  • pip install laya; laya-serve exposes /v1/systemone.
  • 512-token context; multilingual and typed-decisions checkpoints.

Pricing

  • Free open weights (Apache-2.0); you pay your own compute. [Verified 2026-10-15 on Hugging Face model card]

Limits

  • 512-token context. [Verified 2026-10-15 on Hugging Face model card]

Status

  • Model card and GitHub repo live. Release date 2026-09-18 per a third-party site. [Verified 2026-10-15 (repo live); release date unverified]
  • Our smoke test (script: /workspace/research/bench/laya_bench.py): local CPU on an 8-vCPU box, median 181 ms and p95 245 ms per call, 10/10 correct on easy synthetic items, stable across 5 repeats and reversed option order. Small smoke test only, not a quality benchmark. Confidences are uncalibrated (the loader warns the checkpoint ships invalid temperatures). [Verified 2026-10-15 — AgentAPIList local run]

Safety & fit notes

  • Tiny and fast enough for CPU routing and gating; recalibrate before using confidence thresholds.
  • Short 512-token context; truncate or summarize long state.

Sources

Related