Nace AI 发布 Drex 开源决策模型系列，27B 版逼近 Claude Opus 5.5
Nace AI 开源了 Drex 决策模型，27B 版能直接输出选项概率、183 毫秒出结果，成绩快追上 Claude Opus 5.5，做分类或路由的可以试试。
Nace AI 发布 Drex 开源决策模型系列，包含 8B、9B、35B（3B 激活）和 27B 四个规格。模型接收状态输入（文档、记录或智能体上下文）加固定选项，直接输出各选项概率而不生成文本。27B 的 Max 版在 Decision Index 0.4.1 上得 67.52 分，高于 Jev 1.14.0 的 60.70，在 Decision Bench v4 上与 Claude Opus 5.5 相差 1.6 分，与 GPT-6 Sol 和 d3-turbo 统计上打平，响应时间 183 毫秒，远快于 3.5 秒。训练上，模型在微调后加入 RLAF（基于调整反馈的强化学习），贡献 2.0 个 Decision Index 点，配合一致性损失把选项重排导致的答案翻转率从 6.3% 降到 4.2%。
Banger report on foundation decision models. This work presents a 27B open decision model that lands within 1.6 points of Claude Opus 5.5 on Decision Bench v4 and answers in 183 ms instead of 3.5s. Nace AI released Drex, a family of four open decision models at 8B, 9B, 35B with 3B active, and 27B. You give the model a state, such as a document, a record, or an agent's context, plus questions with a fixed set of options. It returns a probability for each option without generating text. The 27B Max model scores 67.52 on the public Decision Index 0.4.1 suite, against 60.70 for Jev 1.14.0. On Decision Bench v4, it is statistically tied with GPT-6 Sol and d3-turbo. It also has the best accuracy among decision models on DecideBench. Model Training: After fine-tuning, training adds reinforcement learning from adjusted feedback (RLAF). A teacher model proposes a probability for each option, and the verified answer adjusts it, so the model trains on calibrated targets in place of one-hot labels. RLAF adds 2.0 Decision Index points, and a consistency loss lowers the rate of flipped answers when the options are reordered from 6.3% to 4.2%. HF: huggingface.co/nace-ai/drex-m… Paper: arxiv.org/abs/2611.07854 Chat with Paper: academy.dair.ai/papers/towards… 💬 17 🔄 3 ❤️ 20 👀 2611 📊 19 ⚡
