Laya：开源 System 1 决策引擎（对照 Jev） Laya: open-source System 1 decision engine (Jev-adjacent)

来源Sources

为什么推荐Why we recommend it 收录理由：与站内 TypeSafe Jev 同类的 System One / 非自回归结构化决策模型，但权重开源（HF 模型页 2026-09-19 核到 license: apache-2.0），HN 当日有讨论，适合 Models 看板对照阅读。核实边界：作者长文对「谁先提出」与 TypeSafe Jev 存在争议，本库不站队、只记双方公开页可达；延迟、基准、与 Jev 的对比表数字全部为作者自述或作者引用的第三方整理（AbdelStark、nibzard 等），本库未复跑，亦未实测 pip 包或 HF Space。五条来源 URL（HF 模型页、作者长文页、HN、GitHub 仓库、HF Space）于 2026-09-19 均核到 200。 Why: a System One / non-autoregressive structured-decision model in the same class as the catalogued TypeSafe Jev, but with open weights (the HF model page showed license: apache-2.0 on 2026-09-19) and a same-day HN thread — useful as a side-by-side read on the Models board. Boundary: the author’s long-form page disputes “who proposed it first” with TypeSafe Jev; this library takes no side and only records that both public pages are reachable. All latency, benchmark, and Jev-comparison figures are author claims or third-party write-ups the author cites (AbdelStark, nibzard, etc.) — not re-run here; the pip package and HF Space were not tested. All five source URLs (HF model page, author page, HN, GitHub repo, HF Space) returned 200 on 2026-09-19.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• 作者称：非自回归、基于双向编码器的 System 1 决策模型，单次前向对任意状态（文本、邮件、工单、JSON）输出结构化决策与校准概率，不生成文本。 Author claim: a non-autoregressive System 1 decision model built on bidirectional encoders; one forward pass over any state (text, email, ticket, JSON) returns structured decisions with calibrated probabilities — no text generation.
• 作者称：三类原语 — choice（从候选字典选一项，返回选项、分布与置信度）、score（按序数评分表定级）、noul（布尔问题，返回校准的 P(true)）。 Author claim: three primitives — choice (pick one key from a criteria dictionary, with distribution and confidence), score (place the state on an ordinal rubric), and noul (a boolean question returning calibrated P(true)).
• 作者称：单 GPU 约 32.8ms（亚 35ms），多语路由覆盖 100+ 语言，权重 Apache 2.0 开源、可自托管；页面自述「6–8 倍快于 Jev」及多项对比表数字，本库未复跑。 Author claim: about 32.8ms on a single GPU (sub-35ms), multilingual routing across 100+ languages, Apache 2.0 open weights, self-hostable; the page’s “6–8× faster than Jev” and comparison-table figures are author claims not re-run here.
• 作者自述局限：choice 选项超过 20 个时表现下降；开箱基础模型在 typed-decisions 基准上接近随机，公布分数来自在基准训练集上微调后的结果（作者原文如此）。 Author-stated limits: choice degrades beyond ~20 options; out-of-the-box base models score near random on the typed-decisions benchmark, and the published score comes after fine-tuning on that benchmark’s train split (per the author).