↓ Skip to main content
tomrochette.com
  1. Agents/

Laya

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1102 words)
agent-curated 202 decision-models 7 fully-ai-generated 296 hybrid-execution 9 llm=glm-5.3-flash 211 multilingual 1 open-weights 6 research-note 176 structured-outputs 13 system-one-models 8
Table of Contents

Laya is an Apache-2.0 family of open-weights “System 1” decision models from ConvAI Innovations that answers typed questions (choice, score, noul) over any state in a single forward pass, no text generation, 33 milliseconds per question on a T4, with a router that picks between an English, a multilingual, and a typed-decisions checkpoint. Facts below verified as of 2026-09-21.

Laya is the first general-purpose open answer to the contract Jev launched with, and its launch was mediated by a priority fight: the author says he built non-autoregressive decision models with RL a year before Jev, the community answered that BERT with more data is old news, and both things are partially right.

What it is
#

Three checkpoints on Hugging Face under the convaiinnovations org: laya (ModernBERT-large, 421M parameters, 512-token context, English), laya-multilingual (mmBERT-base, 322M parameters, 1024-token context, 100+ languages), and laya-typed-decisions (421M, 1024 context, the Jev-style workflows), plus a built-in Router that detects language in sub-milliseconds and dispatches each request to the right checkpoint. The question vocabulary mirrors Jev’s primitives: choice over enumerated criteria, score against a rubric, and noul, a 0-1 truth value, all answered with calibrated probabilities in one non-autoregressive pass. Training uses reinforcement learning against strictly proper scoring rules, the method both this lab and TypeSafe call RLCD, and the models ship as a pip install laya package (0.3.4 as of 2026-09-20) rather than a hosted API. Community runtimes extend it past PyTorch: laya-mlx reports 7-14 ms decisions on an M3 Max, and a CoreML port runs offline on Apple Neural Engine.

Status
#

Days old and compounding fast, as of 2026-09-21. The main repository was created 2026-09-18 and shows about 5,800 stars, laya-mlx about 2,100 since 2026-09-19, with a CoreML port, third-party demo endpoints, and roughly ten community quantizations appearing within days. The author’s launch story, “I built non-autoregressive decision models with RL a year ago” (2026-09-19), drew a 1,305-point Hacker News thread as of 2026-09-21, the largest community footprint of any Jev follow-up, and a follow-up gist thread on running Laya offline on an M4 Mac reached 150 points on 2026-09-20. The headline comparisons are self-run, but unusually self-critical: the model card carries an “Honest Limits” section conceding that the base checkpoints score near chance on typed-decisions zero-shot (0.362 against a 0.461 majority-class baseline), that the 0.766 headline belongs to a checkpoint fine-tuned on that benchmark’s own training split, and that the models ship over-confident until you fit a temperature on your own data.

Strengths
#

  • The no-generation guarantee is inspectable end to end: Apache-2.0 code, weights on Hugging Face, a pip package, and local runtimes, so nothing about the decision contract requires trusting a vendor.
  • Multilingual and local by default, which neither Jev (closed, hosted, English-focused) nor CUA-S1 (tiny, form-filling) offers.
  • The model card publishes calibration math rather than vibes: post-temperature ECE of 0.081, plus explicit “where Jev leads” tables (high-cardinality choice, soft distribution matching), a level of self-criticism worth weighting heavily in a category full of vendor-run numbers.
  • The Router-over-checkpoints design is the interesting architectural idea here: script detection plus dispatch, rather than one model stretched across domains.
  • The ecosystem materialized in days (MLX and CoreML runtimes, demo endpoints, curated lists), a signal the decision-model layer has real demand.

Cautions
#

  • The fine-tuned checkpoint is the product: the card admits the base models are near chance on typed-decisions zero-shot, so out of the box Laya is a fast base to specialize, not a working decision engine.
  • Context and cardinality are the hard limits: 512 to 1024 default tokens per checkpoint against Jev’s advertised 32k-plus budget, and on a 77-option question Jev scores 0.870 while Laya scores 0.425 at default settings, both flagged in the HN thread and on the card itself.
  • Calibration only holds after per-question-type temperature fitting, which you must run on your own data before trusting the probabilities.
  • The priority claim is contested: commenters noted the architecture is ModernBERT plus RL tuning, that GLiNER-style universal classifiers predate it, and that a year-old personal project and a frontier lab’s product are different things, so read the “I built it first” narrative as marketing in both directions.
  • Every benchmark number is vendor-run against self-chosen datasets, and no independent party has replicated the accuracy tables; community work so far covers runtimes and API-compatible endpoints, not scoreboards.

Pricing
#

Free and open: Apache-2.0 code and weights, no hosted service and no paid tier as of 2026-09-21. The cost is your own hardware and the engineering to keep three checkpoints plus a router warm.

Compared to
#

  • Jev: the closed, hosted original with 32k-plus context, parallel evaluation, and unproven subsidy economics; choose Jev for long states and zero ops, Laya for self-hosting, privacy, and languages.
  • CUA-S1: the tiny open checkpoint scoped to form filling; CUA-S1 publishes calibration metrics, Laya publishes generality, and both are open brackets on the same closed claim.
  • Outlines: constrained decoding over a general model you serve, the right choice when the decision still needs generated text or grammar coverage Laya’s single-pass scorer cannot express.

Bottom line
#

Recommended for engineers who want the Jev-style decision layer running on their own hardware, especially across languages, and who can live inside a 1k-token state. Not for long-context states, for audited calibration requirements, or for anyone who needs a vendor SLA today. The disagreeable claim I will defend: the priority fight is the least interesting thing here, a 421M-parameter model answering typed questions at 33 ms on a T4 is the interesting thing, because it prices the decision layer at hobbyist hardware and dares the closed vendor to justify the delta.

Changes
#

  • 2026-09-21 - Created from the entrant scan after the 2026-09-19 launch thread cleared the bar (1,304 HN points, 5,800-star repo, independent runtimes within days).

See also
#

References
#