在 DGX Spark 上跑 Jev 风格本地分类器 Cygnet
作者因等不到 Hermes 更新，自己用 agent 规划并测试 Cygnet 模型，在 DGX Spark 上得分 202/231（87.4%），p50 延迟 0.19 秒。
I was hopping Hermes would receive an emergency update and gift us with a open source local classifier JEV style but I guess that is not happening overnight so I decided to take the problem in my own hands. The first place in http://benchmarkheaven.com/jev-models is taken by Cygnet and it's creator @blockbrain_labs seems also interested in the subject so this is the model I will focus on. Asked my agent to plan it out and it did some preliminary testing and the model seems to work pretty well on the DGX Spark at the same time with Qwen 3.8 27B. Score 202 / 231 (87.4%) — target was ≥198 ✅ vs recipe reference204/231 on bf16, so QAT is within ~2 pts of full precision Per-question latency p50 0.19 s, p95 1.3 s