Laya

On a T4 GPU, Laya takes 33 milliseconds per single question, and batch processing can compress that to 7 milliseconds

Small decision models can serve as front-end schedulers or decision nodes in large model systems, playing a key role in Agent workflows and multi-model orchestration architectures.

Laya is an open-source, open-weight classification decision model that can be downloaded and run locally.

The entire Laya model was trained on a single RTX 6000 Pro (96 GB VRAM).

RLCD is a policy-gradient-based reinforcement learning method (an unofficial name coined by the author) that optimizes decision models for strictly proper scoring rules, granting maximum reward only when the model outputs mathematically calibrated probabilities.

Laya pairs a bidirectional ModernBERT-large encoder with a Transformer head trained from scratch that scores [MASK] option tokens.

Pricor's published benchmark uses 29 held-out columns from two distributors: keyword/type heuristics were correct on 24/29 (82.8%), Laya zero-shot on 18/29 (62.1%), fine-tuned Laya on 22/29 (75.9%), and pure prompting of Gemma on 25/29 (86.2%).

Pricor's processing pipeline is: first a rule analyzer finds the header row, row types, column types and number formats, then the fine-tuned multilingual model Laya assigns each column to a closed set of roles, columns Laya is unsure about go to Gemma 4 E4B running locally via Ollama, and anything still uncertain is handed to the user to answer.

Injected text changes Jev's probability output; Laya's score questions and noul tend to follow option labels rather than state in label handling.

Both Jev and Drex support up to 255 options; Jev reaches 79.2% on the full Banking77 test (77 labels); Laya's performance degrades beyond about 20 options under default settings, reaching only 0.425 on Banking77.

On calibration, Jev's ECE in independent tests varies between 0.02 and 0.35 depending on the task; Laya's ECE is 0.466 before temperature fitting (overconfident) and drops to 0.081 after fitting; Drex has not published calibration data.

Context length: Jev is 64K per request and 32K per question, and with the extended-context header it can reach 262K; Laya is only 512 (English)/1024, with multilingual up to 8K.

Language capability: Jev is mainly English-focused, with lower accuracy in other languages including CJK; Drex has not published a language list; Laya covers more than 100 languages, with 45 of 51 tested languages usable, and has a built-in language router.

In zero-shot quality, Jev is the reference benchmark; Drex's vendor reports that on its Decision Index it leads Jev 61.25 to 57.96, but on JevBench it is slightly lower at 85.7% versus Jev's 86.6%; Laya's zero-shot on typed-decisions is only 0.362, below the 0.461 majority-class baseline.

The three System One decision models share the same API: POST /v1/systemone, passing in state and a question mapping; this protocol is defined by TypeSafe, Jev is the first implementation, and Drex and Laya are wire-compatible.

After fine-tuning, Laya reaches 0.766 on the same benchmark, versus the third-party-published Jev figure of 0.727; Drex offers fine-tuning as a vendor service.

Laya uses the Apache 2.0 license, is fully self-hostable (laya-serve), and also has a free Demo Space; it consists of ModernBERT-large 421M and mmBERT-base 322M encoders and can run on a T4 or even a CPU, with CPU latency of 193–464ms.

On the 112 short cases, hosted Jev and PPLX each scored 112, NeoHorse and Clef each scored 104, Flash and Omni each scored 102, Jeff 2B scored 94, Jeff 0.8B scored 77, Laya Typed scored 52, Laya scored 46, and Laya Multilingual scored 25.

SuperLocalMemory performs memory verification through Laya (local Mac side) or Jev (online side) to ensure reliability.

Laya scored 0% on Khmer text yet reported an average confidence of 95.2%, a textbook case of calibration failure.