Nace AI 发布 Drex 技术报告，推出了一系列采用判别式推理的开源决策模型。其中 27B 的 Drex Max 在不生成 token 的情况下，单次前向传递即可返回选项概率，在 Decision Bench v4 上与 GPT-6 Sol 统计打平。该模型单次决策成本仅为 $0.06 且耗时 183 ms，训练方法结合了监督微调与带 KL 惩罚的 RLAF。
This is a brilliant release from @NaceAI .
It goes after the costly part of AI agents, the thousands of small choices, with a 27B model that only picks and never writes.
- Nace AI’s 27B Drex Max statistically ties GPT-6 Sol on Decision Bench v4 without generating a token.
- It reads a context and question, then returns one probability per supplied option in a single forward pass.
- Its RLAF training rewards verified answers plus a reasoning judge’s probabilities, while a KL penalty limits drift from supervised training.
- Drex Max costs $0.06 per 1K decisions at 183 ms, against $10.17 and about 3.5 seconds for Claude Opus 5.5.
Without chain-of-thought or tools, it answers 21% of sampled HLE questions against 56% for Opus.
来源：Rohan Paul · x.com