AgentJev-0.6B, a local open-source agent, beats Laya on Typed Decisions
AgentJev-0.6B scores 79.25% versus Laya's 77.00%, computes 64 candidates in a single forward pass with 92.4% less compute, makes 68ms reflexive decisions without JSON parsing, and ships open weights that run locally.
😱 什么？！仅 0.6B 本地开源决策模型在 Typed Decisions 上压过 Laya ！ AgentJev-0.6B vs Laya： 🔹准确率 79.25% vs 77.00%（+2.25） 🔹64 候选一次前向，KV 前缀复用，计算量砍 92.4%，约 2× 加速 🔹68ms 反射决策，0 输出 token，无 JSON 解析 🔹选项位置无关 + 概率校准，可直接做 routing / risk gate 🔹上下文 2048 vs 1024 • 权重开源，本机跑，不走闭源 API GitHub：https://github.com/malevrigns/agent-jev HF：https://huggingface.co/aimeigaoshou/agent-jev #AgentJev #Laya #Jev #SystemOne #开源AI #本地模型 https://x.com/nft_chen/status/2102297992179200329?s=46
This post quotes another post: “AgentJev-0.6B open-source decision model tops Laya on Typed Decisions”