TypeSafe AI发布了一款名为Jev的专有文本分类模型，声称在决策任务上的性能可与GPT-5.6 Luna相当，但具有更快的速度和更低的成本1。该模型在IMDb电影评论数据集上的Choice API精度达到96.47%，Noul API精度为96.20%，完成评估的总成本约为0.65美元，运行时间约22至23分钟1。
Jev采用基于强化学习校准决策（RLCD）的训练方法1。TypeSafe AI首席执行官表示，该模型的训练数据100%为合成数据1。
截至文章更新时，OpenAI已在2026年DevDay大会上推出Decision API，功能与Jev相似1。此外，已有数百个快速开发的Jev克隆项目问世，但这些项目的性能普遍不及原版1。
TypeSafe AI has released Jev, a proprietary language model designed for text classification that claims to match the performance of GPT-5.6 Luna while operating at significantly lower costs and faster speeds 1. The model employs a training approach called Reinforcement Learning for Calibrated Decisions (RLCD) 1.
On the IMDb movie review dataset, Jev achieved 96.47% accuracy using the Choice API and 96.20% accuracy with the Noul API, completing the evaluation at a total cost of approximately 0.65 dollars within 22 to 23 minutes 1. According to TypeSafe AI's CEO, the model's training data consists entirely of synthetic data 1.
OpenAI introduced a Decision API at the 2026 DevDay conference, offering comparable functionality to Jev 1. Since Jev's debut, numerous rapid-development clones have emerged within the developer community, though their performance generally falls short of the original implementation 1.
评论
还没有评论，欢迎留下第一条。