Microsoft AI says it has finished pre-training MAI-Base-2, a 2.8-trillion-parameter model that the company plans to license through Microsoft Foundry. The announcement came on October 15, 2026, and it leaves the practical questions open. There is no release date, no public endpoint and no price.
That makes MAI-Base-2 a signal about strategy more than a product anyone can try this week. Microsoft is pitching the model against the strongest open-weight systems from China, while keeping its weights behind a licensing arrangement. Here is what the company has said, what it has not, and what the news means for teams that buy or build on AI models.
What Microsoft says it has finished
Microsoft AI says it has completed pre-training for MAI-Base-2, a mixture-of-experts foundation model, and that the model is now in post-training, according to its frontier pre-training post. Post-training is the stage that shapes a base model into something people can use, and Microsoft says it will share those results "soon." Until then, MAI-Base-2 is a base model, not a chatbot or a callable API.
It is the successor to MAI-Base-1, which Microsoft announced in June 2026 with 1 trillion total parameters and 35 billion active. MAI-Thinking-1, built on that earlier model, is already in public preview on Foundry, so Microsoft has a product path that the new model has not reached yet. GIGAZINE reported the completion, and CEO Mustafa Suleyman announced the model on X on October 15.
Inside the 2.8-trillion-parameter design
The headline figure is total parameters, but the number that matters for cost and speed is the active count. In a mixture-of-experts design, the model holds many specialized sub-networks and sends each token through only a few of them. TestMuAI reports that MAI-Base-2 routes each token to 7 of 512 experts plus one shared expert. That is how a model with 2.8 trillion total parameters activates only 80 billion for any given token.
The training data is also worth a close look. Microsoft's pre-training used 50 trillion tokens, with 18 percent vision-language data, and the model's knowledge cutoff is March 2026, according to TokenPost. Microsoft also says the model was not distilled from other models. As Constellation Research reproduces the company's announcement, "Our pre-training is built on organic data which does not involve distillation from other models." Buyers increasingly ask whether a model learned from a rival's outputs, so that sentence is one they should ask vendors to back up in writing.
MAI-Base-2 training run: 38 days on 16,000 GB200 GPUs
The run details are unusually specific for a model without a release schedule. Here are the figures reported across the coverage:
- Training window: August and September 2026
- Hardware: 16,000 NVIDIA GB200 GPUs
- Run length: 38 days, as Microsoft reported when it announced the completion on October 15
- Training data: 50 trillion tokens, including 18 percent vision-language data
- Knowledge cutoff: March 2026
Scale matters beyond bragging rights. Pre-training at this size requires a cluster that few organizations can build, which is why compute supply has become a strategic story. ByteDance is nearing a deal with MediaTek for custom AI data center chips, a reminder that the race is being fought at the silicon level as well as inside model labs.
Microsoft's 4x to 12x efficiency claim against DeepSeek
The most eye-catching claim is comparative. Microsoft says MAI-Base-2 delivers 4x to 12x more pre-training FLOPs efficiency than DeepSeek-V4-Pro-Base, which it describes as the strongest open base model. Both Constellation's coverage and byteiota's review treat the figure as a company claim.
FLOPs efficiency measures how much training compute a model needs to reach a given level of quality. It says nothing about what the model costs to call, how fast it answers, or how it handles real tasks. Until independent evaluators run MAI-Base-2 against DeepSeek's model, the range should be read as Microsoft's own measurement.
Constellation Research notes that the approach sidesteps the open-model label, since Microsoft plans to license weights rather than publish them. That distinction matters for developers who expect open-weight models to be freely downloadable and redistributable. Whether the licensed weights will meet the open-source community's definition of open is unclear.
MAI-Base-2 release date and Foundry licensing
Microsoft has not announced a release date, a partner list or pricing. The plan is to license weights to trusted partners through Foundry. As Constellation Research reproduces the announcement, "This will enable us to license our model weights to trusted partners through Foundry". The sentence describes a route to access, not a public launch.
The enterprise angle is clearer than the consumer one. Microsoft frames MAI-Base-2 as a base for companies to customize with their own data, which makes Foundry customers the most likely early audience. Consumers and small businesses have no way to use the model today.
Not every outlet accepted the completion claim at face value. Archynewsy reported that MAI-Base-2 did not appear on Microsoft's public model list and described the completion as unverified. Microsoft's own post on frontier pre-training names the model and its specifications, so that skepticism looks outdated or incomplete. The core performance claims still rest on Microsoft's word.
How to evaluate model vendors before a model ships
If your team is planning around a model that is not yet available, a few habits will prevent wasted effort.
- Separate base models from post-trained products. A base model is raw material for customization, while the post-trained version is what most teams actually deploy. Microsoft has not said what the post-trained form of MAI-Base-2 will cost or when it will arrive.
- Ask for pricing, data-provenance terms and release dates in writing. Microsoft's no-distillation statement is a useful starting point, but it belongs in contract language, not only in a company post.
- Don't build a roadmap around an unreleased model. Test what you can access now, including MAI-Thinking-1 in public preview on Foundry, and keep at least one alternative vendor in your plan.
- Read the license when weights arrive. Check commercial use, redistribution limits and any restrictions on training or fine-tuning.
What to watch next
Several signals will show whether MAI-Base-2 becomes a usable product. The first is the post-training update Microsoft has promised to share. Next come partner announcements under the Foundry license, Foundry pricing, and any independent benchmark that tests the 4x to 12x claim against DeepSeek-V4-Pro-Base.
The competitive stakes are high. Anthropic reported $16.8 billion in Q3 revenue ahead of a possible IPO, and rival labs face growing pressure to show that expensive training runs produce paying customers. Microsoft's own sequence offers a hint. MAI-Base-1 arrived in June, and MAI-Thinking-1 is already in preview. If that pattern holds, a post-trained product may reach customers before the base weights reach outside teams, though the sources give no timeline.
