Jev launches: a System One model for decisions
TypeSafe founder Diogo Almeida introduces Jev, its RLCD training method, and its focus on software decisions. The post includes a 2:56 launch video.

Nace AI has released Drex Max v1 weights on Hugging Face: 2.6B? Wait, the post says 260亿参数 (26 billion parameters), bf16 about 52GB. The model does not generate text; given a state and questions, it outputs probabilities for each option in one forward pass, suitable for local routing. The model card compares it with Jev 1.13, scoring 67.52.
Nace AI 把 Drex Max v1 的权重挂到 Hugging Face 上了,不用申请就能下,260亿参数,bf16 差不多52GB。这东西不生成文字,给它一段状态和几个是否、单选、打分的问题,一次前向就把每个选项的概率给出来,这种活本来就适合放在本地做分流,Jev 再便宜也只能走 API。模型卡里拿它跟 Jev 1.13 比,67.52