Microsoft AI said Thursday, October 15, it has finished pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active at a time, in a technical post that reads like a declaration of independence written in FLOPs.
The model is still in post-training, the stage that turns a raw base model into something you can chat with. Microsoft gave no release date, API or pricing, Investing.com noted.
Then, at an employee town hall the same day, Microsoft AI CEO Mustafa Suleyman said the quiet part out loud: his team’s base model is stronger than the one powering OpenAI’s GPT-5.6 Sol, and it took a quarter of the compute.
The spec sheet
Per Microsoft’s post and StreetInsider’s summary:
- Size: 2.8 trillion total parameters, 80 billion active per token.
- Compute: 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026.
- Data: a roughly 50 trillion-token training mixture with a March 2026 knowledge cutoff; unique tokens grew about 36% over MAI-Base-1.
- Vision: trained natively with images using an encoder-free design, with 4.5 trillion vision-language tokens, 18% of the mix.
- Efficiency: average model FLOPs utilization of 21% and goodput of 89%.
A quick decoder. “Mixture of experts” means the model is split into many specialist sub-networks, and only a few switch on for each word, so a 2.8 trillion-parameter model runs with the cost profile of an 80 billion-parameter one. “Goodput” is the share of training time that actually produced useful progress rather than being lost to crashes and restarts.
That 21% MFU (model FLOPs utilization) figure is refreshingly honest. It means most of the theoretical horsepower of those 16,000 GPUs went unused, which is normal at this scale but rarely stated so plainly.
The claims, public and private
Microsoft’s post says MAI-Base-2 “outperforms every open-weight model we tested on our held-out evaluations” and is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which it calls the strongest open base model. Investing.com put the range at 4.2x to 12.3x across four evaluations.
The town hall went further. “We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol,” Suleyman told employees, according to Business Insider. “And in fact, we used one-quarter of the compute that OpenAI used to train their model.”
He said Microsoft had tested MAI-Base-2 “privately and confidentially” against OpenAI models, including one nicknamed “Spud” that powers GPT-5.6, and that the comparisons rely on technical information Microsoft has about OpenAI’s models. He also said the model used “12 times less computation than was used for DeepSeek.”
The catch: nobody else can check
Every number above comes from Microsoft. The public claims compare a base model against other base models on Microsoft’s own held-out evaluations, and the OpenAI comparison rests on private data that no outsider can see.
Base-model quality also isn’t the same as product quality. Much of what makes a chatbot useful, from following instructions to refusing the right things to using tools, comes from post-training, the step MAI-Base-2 hasn’t finished. A stronger foundation helps. It doesn’t guarantee a better house.
And there’s no comparison against the closed frontier leaders at the post-trained level, which is the league Microsoft actually wants to play in. Until there’s an API and independent benchmarks, treat this as a credible progress report, not a scoreboard.
Why Microsoft is doing this
Microsoft stresses that its pre-training uses “organic data which does not involve distillation from other models.” That’s a pointed line in an industry where training on rivals’ outputs is a recurring accusation. Microsoft says clean data provenance will let it license the model weights to “trusted partners” through Foundry, its enterprise AI platform.
Our read: the subtext is OpenAI. Microsoft has invested billions in OpenAI and, after OpenAI’s 2025 recapitalization, held a stake valued at about $135 billion, roughly 27%, according to the companies’ October 2025 agreement. That same agreement loosened the ties: Microsoft “can now independently pursue AGI alone or in partnership with third parties,” and it gave up its right of first refusal to be OpenAI’s compute provider.
Suleyman leads what Microsoft calls its superintelligence team, and in March the company reassigned him from running Copilot to focus on building models, Business Insider reported. MAI-Base-2 is the first big public receipt from that shift.
It’s notable that Suleyman would knock OpenAI’s models even in an internal meeting. As Business Insider points out, OpenAI is one of Microsoft’s biggest customers and partners, and criticism risks denting the value of Microsoft’s own stake. That he did it anyway tells you how much Microsoft wants its staff, and eventually its customers, to believe it can stand on its own.
Who it affects
- Copilot and Azure customers: in time, Microsoft could route more traffic to its own models, which would change cost and behavior under the hood.
- OpenAI: a major partner is now openly a competitor at the model layer, not just a distributor.
- Enterprises worried about provenance: a “no distillation” pitch plus licensable weights could appeal to legal and compliance teams.
Microsoft says preparations for MAI-Base-3 are already underway.
The takeaway
Our read: Microsoft now has a homegrown model at frontier-ish scale, a credible efficiency story and a data-provenance pitch enterprise buyers will like. But “our base model beats theirs, trust us” is a town-hall line, not a benchmark.
What to watch: the post-trained version and its API, independent benchmark results, pricing in Foundry, and whether Microsoft starts routing Copilot traffic to its own models instead of OpenAI’s.
Image: Christopher Wilson / Wikimedia Commons, CC BY-SA 4.0
