Wait Which Model?
← Back to directory
Convai Innovations

Laya

Released Sep 18, 2026 · knowledge cutoff unpublished

Status
Unknown
Location
India
Type
System One
Modality
Text
Context window
1K
Max output
—
Speed
—
Price ($/MTok in / out)
— / —
Cost per task
—
Open weights
Yes
Benchmarks0 of 10 reported

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Answers a routed question in about 33 ms on a single T4 and 7 ms per question batched, so the whole decision layer fits on a free Colab GPU or an 808 MB download with no API bill
  • Weights, training recipe and datasets are open, so calibration can be refit on your own labels — the out-of-the-box expected calibration error of 0.466 drops to 0.081 after per-question temperature fitting
  • Speaks the same POST /v1/systemone wire protocol as Jev through laya.serve, so a client written against TypeSafe moves over by changing the base URL and model id
  • A built-in router detects script and language before inference and dispatches to the English or multilingual checkpoint, so mixed-language queues need no branching in the caller

Weaknesses

  • Near chance zero-shot on typed decisions — 0.362 against a 0.318 random floor and a 0.461 majority baseline — so the 0.766 headline is a fine-tuned checkpoint, and Convai's own card calls the base a fast model to specialise rather than a zero-shot decision engine
  • Accuracy falls off sharply past roughly 20 options because each label gets only three or four tokens of the 256-token question budget — a 77-way Banking77 split scored 0.425 where Jev scored 0.870
  • Confidence does not track ability off-distribution: the English checkpoint scored 0.000 on Khmer while reporting 0.952 confidence, an independent adversarial set of 100 hand-written English scenarios came in at 40.3% zero-shot, and noul answers can follow the option labels instead of the state
  • Behavior is a moving target — eight PyPI releases in its first six days, a quickstart whose documented outputs no longer match the current release, and CPU-only runs that one reviewer clocked at 49 s per prediction against the 193-464 ms Convai publishes

Not an LLM: an open-weights "System One" decision model in the same mould as TypeSafe's Jev — a bidirectional encoder (ModernBERT-large, 421M) with a decision head trained by RLCD against strictly proper scoring rules that takes text state plus typed questions (choice, score, noul) and returns calibrated probabilities in one forward pass, so none of this site's benchmarks apply and every cell is null by design. Filed under the fast tier alongside Jev. Context is 512 tokens on the root convaiinnovations/laya checkpoint (about 320 for state); the laya-multilingual (mmBERT-base, 322M, 100+ languages) and laya-typed-decisions checkpoints take 1,024, the mmBERT encoder can be configured to 8,192, and impossibl's hosted endpoint is widened to 8,192. Convai publishes no API or price, so pricing is null; availability is general because impossibl serves both checkpoints free at api.impossibl.com/v1/systemone (documented in Laya's README from 2026-09-23) and the weights run on a free Colab T4, with a demo Space on Hugging Face. Released as laya 0.1.0 on PyPI and Hugging Face on 2026-09-18, three days after Jev, and at 0.3.7 by 2026-09-23. Convai's Jev comparisons (0.766 vs 0.727 on its typed-decisions set, 32.8 ms vs 236-276 ms, ECE 0.081 vs 0.246) are its own measurements and set a local T4 forward pass against Jev's hosted-API round trip; the zero-shot, Khmer, Banking77, adversarial and CPU figures in weaknesses are from the model card and independent tests (orcarouter, flowtivity, eesel, invidelabs). Founder Nandakishor Mukkunnoth frames Laya as a reply to TypeSafe presenting the concept as new, citing his own arXiv 2503.23303 (March 2025) and 2510.01237 as prior art. No Artificial Analysis coverage; knowledge cutoff unpublished.

For developers

API model strings

  • impossiblconvaiinnovations/laya
  • impossiblconvaiinnovations/laya-multilingual

Licence

Apache License 2.0 · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News