AI 热点

微软发布 MAI Cyber 2 Flash、MAI Code 2 等七款 MAI 模型

Mustafa Suleyman更新于 11月18日 02:34

Secure, aligned & contained frontier intelligence at a fraction of today’s cost

AI 导读

微软发布七款 MAI 模型,其中 MAI Cyber 2 Flash 在 CyberGymE2E 上超过 GPT 6 Astra,单任务成本不到其 2%;MAI Code 2 在日常编码任务上达到 GPT-6.1 Sol 98% 的性能,价格为其 40%。编码评测采用 DeepSWE 1.1、Terminal Bench 2.1、SWE-Atlas 和 FrontierCode v1.1 的平均值,同时推出 MAI Licensed Weights,支持企业用自有数据定制模型并在自有 Azure 环境运行。

推荐理由

精选AI 评分 86

微软一次性发布七款 MAI 模型,并给出与竞品的价格性能对比,可据此判断低成本专用模型在编码与安全场景的定位。

主题

标签

正文

Today we're launching MAI Cyber 2 Flash which beats GPT 6 Astra on CyberGymE2E at just 2% of the cost per task allowing defenders to run it continuously in the background.

Similarly, MAI Code 2 gets you 98% of the performance of GPT-6.1 Sol on everyday coding tasks to help with patching and remediation, at just 40% of its price.

The last few months have been a warning shot for everyone. It's very clear that attackers are gaining momentum, as exploits become cheaper and easier than ever.

Optimizing the price-performance frontier is critical if we're to enable developers to secure their code, systems, and networks against vulnerabilities.

MAI Code 2 gives you 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at just 5% of the cost. For coding, we use an average of DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 as basket of practical, everyday coding benchmarks that best reflect the usage we see in GitHub Copilot.

MAI Code 2 Flash matches GPT-6 Luna and Claude Haiku 5.5 on those same everyday tasks, at 65% and 20% of the price.

MAI Cyber 2 Flash matches GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGymE2E benchmark.

It’s important to say, we’ve built all these models to prioritize the coding work our customers do every day. With these Code and Cyber lineages we aren't chasing niche capabilities at your expense. Because when the same model is built to also push the frontier of science, you pay for that capability on every task, even for a routine bug fix.

We also have some of the very best multimodal models in the world.

MAI Image 2.7 is #1 on Artificial Analysis image editing, outperforming GPT Image 2.6 Sunburst (max) at just 20% of the price.

Dragon Copilot Transcribe is built on MAI-Transcribe-2, the world's fastest and most accurate multilingual transcription model. It already powers 10 million patient encounters a month – halving multilingual error rates.

MAI Realtime 1 outperforms GPT-Realtime 2.1 and starts at just $0.01 per audio minute.

Decision 1.1 the latest decision model from Microsoft, which integrates right into your applications. It beats Jev, Google and OpenAI on the AA Decision Index.

MAI Licensed Weights I’ve also heard again and again that enterprises don’t want to pour all their valuable IP, data and tacit knowledge into someone else’s model. They want control of their own weights.

With MAI Licensed Weights, we’re offering customization and control. Take a frontier MAI model, customize it using your proprietary data. Then run it securely in your own Azure environment.

You control who can access it. Your data and adaptations stay yours. And we’ll work with you to deliver custom models and embed them in your own products and agents.

Seven new MAI models! The hill climbing machine is motoring!

来源

Mustafa Suleyman