aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

Nace AI releases 27B decision model Drex Max to score supplied optionsMachine translation

AI introduction

Drex Max reads a context and question, then returns a probability for each supplied option in one forward pass without generating answer text. The post says it statistically ties GPT-6 Sol on Decision Bench v4 and costs $0.06 per 1,000 decisions at 183 ms. Without chain-of-thought or tools, it answers 21% of sampled HLE questions.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

@NaceAI 的这次发布非常出色。

它用一个只做选择、从不生成文本的 27B 模型,瞄准了 AI 智能体成本高昂的环节:成千上万个细小决策。

- Nace AI 的 27B Drex Max 在 Decision Bench v4 上与 GPT-6 Sol 在统计意义上打成平手,而且无需生成任何 token。

- 它读取上下文和问题,然后通过一次前向传播,为每个给定选项返回一个概率。

- 它的 RLAF 训练会对经过验证的答案以及推理评判模型给出的概率给予奖励,同时用 KL 惩罚限制模型偏离监督训练的结果。

- Drex Max 每 1K 次决策的成本为 0.06 美元,耗时 183 毫秒;Claude Opus 5.5 则为 10.17 美元和约 3.5 秒。

在没有思维链或工具的情况下,它答对了抽样 HLE 问题中的 21%,而 Opus 为 56%。

QuoteNace AI@NaceAI
Translation
决策模型是如何训练的? 今天,我们面向开发者社区发布 Drex 技术报告!🚀 了解这些先进的开源决策模型如何通过 RLAF 训练的详情。 访问 https://t.co/swZ0gsqCyo,构建更多应用(免费试用)。 附言:Drex Max 1.1 和 Drex for Finance 是闭源产品,专为企业用例设计。
Original

How are Decision Models Trained? Today, we are publishing the Drex Technical Report for the developer community! 🚀 Discover details on how the State-of-the-art Open-Source decision models are trained with RLAF. Build more at: https://t.co/swZ0gsqCyo (try for free) P.S. Drex Max 1.1 and Drex for Finance are closed sourced and designed for enterprise use-cases

View this post on X ↗

Source: Rohan Paul · x.com

Model
Found an error? Send a correction