Laya
An open-weights System One model from Convai Innovations. A fully fine-tuned ModernBERT-large encoder with a from-scratch decision head that scores one marker per option and answers every question in a single 33–39 ms pass. Runs on your own CPU or GPU.
Laya is trained with Convai's RLCD recipe and proper scoring rules; the base is stated in the README. The repository carries three checkpoints: the English base (421M, 512-token context), a multilingual checkpoint (322M on mmBERT, 1,024 tokens, extendable to 8,192) and a typed-decisions checkpoint. On the typed-decisions suite the base checkpoint scores near chance (36.2%) while the typed-decisions checkpoint scores 76.6%, so pick the checkpoint for the job. Reported calibration error is 0.081 after temperature fitting (0.466 as shipped). laya-serve exposes /v1/systemone; community ports exist for Core ML, MLX, GGUF, ONNX and LiteRT. Fine-tune it on a Mac with Laya Studio.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 421M |
| Base model | answerdotai/modernbert-large |
| Maker | Convai Innovations |
| Released | 2026-09-18 |
| License | apache-2.0 |
| Reported latency | 39.5 ms per question on a T4 (English); 32.8 ms multilingual; 193–464 ms on CPU |
Get the weights
pip install systemonemodels
systemone pull convai-innovations/laya
The files are served from the maker's Hugging Face repository, convaiinnovations/laya, and verified against the checksums recorded here.
Read more
This page was opened by System One for Convai Innovations, who can claim the organisation and take it over at any time.