Microsoft rolls out seven homegrown AI models and a weights licensing plan
At its Ignite conference, Microsoft unveiled seven in-house AI models spanning coding, cybersecurity, image, voice, transcription and decision-making — pitched as matching top rivals at lower cost. The company also opened a licensing program giving selected enterprises direct access to its model weights, all of which must run on Azure.
- Original source: heise.de.
- The source published it on Nov 18, 2026 at 7:39 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Microsoft used its Ignite conference to lift the lid on seven of its own AI models, a lineup that puts it squarely up against OpenAI and Anthropic. The set covers coding, cybersecurity, image generation, voice conversations, transcription and decision-making. Microsoft says each is built to match the leaders in its category while costing less to run.
The coding flagship is MAI Code 2, which Microsoft says averages 98% of GPT-6.1 Sol's performance across four benchmarks at roughly 20% lower cost without an action discount. A faster sibling, MAI Code 2 Flash, targets routine work. Both are already plugged into GitHub Copilot.
On security, Microsoft claims MAI Cyber 2 Flash performs comparably to GPT-6 Astra at finding vulnerabilities, with a calculated cost per test task more than 98% lower. Three more models round out the low-cost push: MAI Image 2.7, MAI Realtime 1 and MAI Transcribe 2. The decision-making model Microsoft Decision 1.1 is already live in Copilot, routing user requests between Copilot itself and the AI agent Cowork.
Microsoft isn't stopping at its own products. Under a new licensing program, selected companies will get direct access to the weights of its AI models so they can tailor them and run them inside their own Azure environments. There is no public release of those weights. Microsoft says it is already building customized models with AT&T, Toyota and Alaska Airlines.
In his Ignite keynote, AI chief Mustafa Suleyman said customers can contribute their own expertise without surrendering control of their data and customizations. Because the models have to run on Azure, the arrangement could also feed Microsoft's cloud business — and heavier use of in-house models would loosen its reliance on OpenAI and other developers.
Longer term, Microsoft wants a seat at the general-purpose flagship table. Suleyman told Axios the company is making significant progress on such a model, with top-tier results already in pretraining and reinforcement learning work advancing. He gave no release date.
TopicsMicrosoft · OpenAI · Anthropic · Mustafa Suleyman · Azure · AT&T · Toyota · Alaska Airlines
Source: heise.de
Written by AI from verified reporting.

