WHAT NEW IN X10.29 THU

AI collections from X · 22 collections · 58,925 posts

jevIMAGE

AgentJev-0.6B open-source decision model tops Laya on Typed Decisions

Built on Qwen3-0.6B, this open System-1 decision model runs Boolean, Choice and Score with zero decode tokens and reaches 79.25% Top-1 on 2000 questions, with full deployment steps included.

SuSu_酥酥👅
@NFT_ChenArchived 2026.09.23Originally posted 2026.09.22
AgentJev-0.6B open-source decision model tops Laya on Typed Decisions (image 1 of 3)
AgentJev-0.6B open-source decision model tops Laya on Typed Decisions (image 2 of 3)
AgentJev-0.6B open-source decision model tops Laya on Typed Decisions (image 3 of 3)
SuSu_酥酥👅 @NFT_ChenOriginally posted 2026.09.22
Open original on X

🤯卧槽!开源圈又杀出个比 Laya 更狠的 System-1 决策核模型 AgentJev-0.6B ! Typed Decisions 2000 题 Top-1:79.25% vs Laya 77.00%,硬超 +2.25pt。 底座 Qwen3-0.6B,Boolean / Choice / Score 全开,解码 token = 0。 比 Laya 更狠的点: 🔹准确率更高:1585/2000,纸面领先 2.25 个百分点 🔹宽候选更快:64 个动作时 Shared Prefix KV Cache,骨干 token 从 33,547 砍到 2,551(-92.4%),约 2× 🔹状态吃得更长:上下文 2048 vs Laya 1024 🔹输出可当闸门:完整概率分布 + 分类型校准,直接做 routing / threshold / fallback 🔹权重、训练、推理全开源,能复现、能继续训 上手步骤: 1️⃣git clone https://github.com/malevrigns/agent-jev.git,建 venv,装 requirements.txt 2️⃣从 HF 拉 aimeigaoshou/agent-jev 的 model.safetensors + temperatures.json 3️⃣把 safetensors 包成 agentjev_v1.pt(不能当普通 CausalLM 加载) 4️⃣起服务:python -m jev_service.server --checkpoint agentjev_v1.pt --model-path Qwen/Qwen3-0.6B --temperatures temperatures.json --port 8149 5️⃣用 agentjev_client.py 丢 state,调 decide_boolean / decide_choice / Score,拿分布做门控 目前只打赢 decision benchmark,真实 Agent 轨迹还没完整对打。 
HF:https://huggingface.co/aimeigaoshou/agent-jev #AgentJev #Laya #Jev #TypedDecisions #AIAgent #开源 #Qwen3 #SystemOne #KVCache #决策模型

This post quotes another post: x.com/nft_chen/status/2102277701415600236

x.com/nft_chen/status/2102297992179200329Archived 2026.09.23 10:18