微软加入了Jev粉丝俱乐部，这是一个由公司组成的意外群体，他们共同希望因各自的决策模型而得到认可。Jev由TypeSafe AI在三周前宣布，是一种经过调优的大语言模型（LLM），能够以有限的响应范围回答特定类型的问题，并按概率进行评分。由于其速度快、相对成本低以及响应受限，它非常适合于那些不希望出现开放式文本且准确性不确定的各种商业应用场景。
Microsoft has joined the Jev fan club, an accidental group of companies that share a common desire to be recognized for their own decision models. Jev, announced by TypeSafe AI three weeks ago, is a large language model (LLM) tuned to respond to certain types of questions with a limited range of responses, rated by probability. Due to its speed, relative affordability, and response constraints, it's well-suited for a variety of business applications where open-ended text of uncertain accuracy might be undesirable.
Jev的一个卖点是，其决策模型不会像标准LLM那样产生幻觉。但决策模型仍可能出错，且其受欢迎程度已经促使研究人员探讨这些错误可能如何被放大。尽管如此，它们仍有其用途。对Jev的关注促使其他公司宣布他们也有决策模型可供提供，尽管机器学习研究者早已能够创建用于基于概率决策的分类器模型。OpenAI表示其Decisions API已进入公开测试阶段。
One of the selling points of Jev is that decision models don't hallucinate in the way that standard LLMs do. But decision models can make errors and their popularity has already prompted researchers to explore how those errors might be magnified. Nonetheless, they have their uses. The attention lavished on Jev prompted other companies to declare that they too have decision models to offer, even though machine learning researchers have long been able to create classifier models for probability-based decisions. OpenAI said its Decisions API has entered public beta.
Cloudflare也加入了讨论，推出了其Clef模型。Strands推出了Strands Decider 2B，称其为“一个小型、开源的决策模型”。Liquid AI推出了d1。Perplexity推出了其Decisions API。Snowflake大力宣传了其自身的决策模型。还有Surogate Rune和H2O.ai的H2O-Lightning-4B，这是一个基于Qwen3.5-4B的决策模型。总而言之，目前已有超过100种此类模型在竞争关注度。
Cloudflare chimed in with its Clef model. Strands trotted out Strands Decider 2B, "a small, open source, decision model." Liquid AI introduced d1. Perplexity launched its Decisions API. Snowflake talked up its own decision model. Then there's Surogate Rune and H2O.ai's H2O-Lightning-4B, a decision model built on Qwen3.5-4B. All told, more than 100 such models are now vying for attention.
现在轮到微软了。“决策模型正迅速成为AI领域的一个重要新类别，”微软CTO办公室软件工程副总裁Achint Srivastava在周五的一篇博客文章中表示。“与设计用于生成文本或通过复杂问题进行推理的LLM不同，决策模型是专门构建用于提供软件可以立即采取行动的结构化输出的。”
Now it's Microsoft's turn. "Decision models are quickly emerging as an important new category in AI," said Achint Srivastava, VP of software engineering in the Office of the CTO at Microsoft, in a blog post on Friday. "Unlike LLMs, which are designed to generate text or reason through complex problems, decision models are purpose-built to deliver structured outputs that software can immediately act on.
一旦你理解了这种能力——以高性能以极低的成本进行决策和分类——各种有用的任务就会被解锁。微软参与这场竞赛的模型名为Microsoft-Decision-1，将通过微软Foundry提供，不久后还将通过OpenRouter提供。雷德蒙的决策模型与H2O的模型类似，均基于Qwen模型，此次采用的是Qwen3.5-9B。Qwen模型系列由阿里巴巴云开发，阿里巴巴云是中国科技巨头阿里巴巴的云计算部门。
And once you understand that capability – making decisions and classifying things at very low cost with high performance – all kinds of useful tasks get unlocked." Microsoft's entrant into the race is called Microsoft-Decision-1, which is offered via Microsoft Foundry and, soon, via OpenRouter. Redmond's decision model, like H2O's, is based on a Qwen model, Qwen3.5-9B in this instance. The Qwen model family is developed by Alibaba Cloud, the cloud computing arm of Chinese tech giant Alibaba.
微软因未披露的原因表示，它将很快将Decision-1重新基于其自身模型以及来自OpenAI的模型。这家美国软件巨头声称，其模型在延迟测试中比H2O-Lightning-4B快2.5倍，比Jev快2.8倍；在36个基准测试中的准确率领先群雄（83.5%）；在置信度得分上排名第二（92.2%），仅次于Quyet-1.0-Large；在文本分类任务中比OpenAI的GPT-6 Sol便宜20倍以上。
Microsoft, for reasons not disclosed, said it will soon rebase the Decision-1 on its models and those from OpenAI. The US software giant claims its model is 2.5x faster than H2O-Lightning-4B and 2.8x faster than Jev in its latency test, leads the pack in accuracy (83.5 percent) on 36 benchmarks, ranks second (behind Quyet-1.0-Large) in confidence score (92.2 percent), and is more than 20x cheaper than OpenAI's GPT-6 Sol in text classification tasks.
输入token的成本为每百万token 0.042美元，输出token免费。斯里瓦斯塔瓦表示：“现在，代理AI已经成为现实，我们看到成本在人们决定如何使用AI方面起着重要作用。” “并且，为正确的工作选择正确的模型变得越来越重要。”®
Input tokens cost $0.042 per million tokens and output tokens are free. "Now that agentic AI is a reality, we’ve seen that cost plays a major role in how people decide to use AI," said Srivastava. "And it’s increasingly important to choose the right model for the right job." ®