WHAT NEW IN X10.31 SAT

AI collections from X · 22 collections · 62,376 posts

jevIMAGE

Open-source Laya decision model claims to beat Jev with 33ms results

The 421M non-autoregressive Laya model claims higher typed-decision accuracy and better calibration than Jev, with pip install and predict usage examples.

SuSu_酥酥👅
@NFT_ChenArchived 2026.09.20Originally posted 2026.09.19
Open-source Laya decision model claims to beat Jev with 33ms results (image 1 of 3)
Open-source Laya decision model claims to beat Jev with 33ms results (image 2 of 3)
Open-source Laya decision model claims to beat Jev with 33ms results (image 3 of 3)
SuSu_酥酥👅 @NFT_ChenOriginally posted 2026.09.19
Open original on X

🔥速码:Jev 要排队、要付费、还闭源!开源加强版 Laya 已经能在消费级 GPU 上把它打穿! 421M 非自回归决策模型,一次前向传播出结果,不写字、不幻觉、不解析 JSON。 核心亮点: 🔹单题约 33ms,比 Jev 快 6–10 倍 🔹typed-decisions 准确率 0.766 > Jev 0.727 🔹校准更好:ECE 0.081 vs 0.246 🔹支持 choice / score / noul 三种判断 🔹100+ 语言,Apache 2.0,本地跑 $0 🔹50% 自动放行时准确率约 92% 操作步骤: 1️⃣pip install laya 2️⃣agent = laya.load("convaiinnovations/laya") 3️⃣agent.predict(state, questions)
把邮件/工单丢进去,同时问:该分给谁、多紧急、会不会流失。一次返回带概率的答案。 生产建议用 laya-typed-decisions,多语言走 Router()。 模型:http://huggingface.co/convaiinnovations/laya 演示:http://huggingface.co/spaces/convaiinnovations/laya-demo #Laya #Jev #SystemOne #开源AI #决策模型 #LocalLLM #HuggingFace #RLCD #加强版

This post quotes another post: x.com/nft_chen/status/2101253568774697099

x.com/nft_chen/status/2101268965167399143Archived 2026.09.20 10:15