JEVfire：不改模型让 Qwen 用 Jev 式决策，快 8–14 倍
给 Qwen 固定候选答案菜单、读分数、用代码拼 JSON，28 次决策从 5.11 秒降到 497 毫秒，还在浏览器里跑本地游戏决策。
Someone just made ordinary Qwen behave a LOT more like Jev without training a new model. 🏆 TOP Jev Clone (according to HF leaderboard) 🎯JEVfire takes an existing open LLM and changes how it makes decisions. Instead of prompt → generate { → key → : → value → comma → … It gives Qwen a fixed menu of legal answers, reads the model’s scores, picks the winners, and builds the JSON in code. No retraining. No second model. No autoregressive JSON serialization. On Qwen3.8-27B-FP8 + vLLM ... 4 decisions 🐌 JSON generation: 878 ms ⚡ JEVfire: 110 ms 🚀 8× faster 28 decisions: 🐌 JSON: 5.11 sec ⚡ JEVfire: 497 ms 10.3× faster With a cached prefix ... 🚀 346 ms 👉 14.8× faster They put Qwen3.5-0.8B in a browser with WebLLM/WebGPU and had it make real-time game decisions locally. 🍄 Mario-style demo 🧠 ~450MB model 💾 ~2GB free GPU/unified memory ⚡ 71 ms/model decision on an M4 Max ☁️ no inference server 🔑 no API key This is NOT TypeSafe’s Jev and it doesn’t reproduce Jev’s RLCD training. But the idea is fascinating! Maybe making Local AI agents fast isn’t only about smaller models. 🔗 Link in ALT