Microsoft's MAI-Base-2 Finishes Pre-Training at 2.8T Parameters

Microsoft AI has finished pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters, and says its new training recipe is between 4.2 and 12.3 times more compute-efficient than the strongest open base model it could benchmark against. According to the company's Lab Notes post "Frontier Pre-Training at Microsoft AI," published October 15, 2026, the model activates only 80 billion parameters per token and was trained across 38 days in August and September on 16,000 Nvidia GB200 GPUs. The announcement marks the largest publicly confirmed model in Microsoft's own AI family — and comes one day after Microsoft AI CEO Mustafa Suleyman told employees the new model outperformed the base model behind OpenAI's GPT-5.6 Sol using a quarter of the compute.

What Microsoft built: 2.8 trillion parameters on 16,000 GB200s

MAI-Base-2 holds 2.8 trillion total parameters but routes each input through only 80 billion of them — an activation ratio of about 2.9 percent that is the core efficiency mechanism of its mixture-of-experts design. According to reporting on the Lab Notes post, the training consumed a 50-trillion-token mixture with a March 2026 knowledge cutoff, and unlike its predecessor, MAI-Base-2 trained on vision from the start: roughly 18 percent of the data combined images and text, covering charts, captions, spatial reasoning and medical images. Microsoft reports 21 percent average model FLOPs utilization and 89 percent goodput across the run, meaning the cluster made useful training progress nearly nine-tenths of the time — a strong figure for a training effort of this scale.

Perhaps the most consequential line in the technical disclosure is about data provenance rather than scale. According to an analysis of the announcement, Microsoft emphasizes that the model was built entirely from organic data with no distillation from other companies' model outputs — a deliberate choice that determines what Microsoft can legally do with the weights. Many rival open models are trained partly on synthetic output from other labs' models, which creates ambiguity about who owns the resulting weights; organic data sidesteps that, which is why Microsoft says it will be able to license MAI-Base-2's weights to trusted partners through its Foundry platform rather than offering access only through an API.

The 12x efficiency claim — and how to read it

The headline efficiency number deserves careful reading. Microsoft is not claiming MAI-Base-2 beats DeepSeek in a benchmark head-to-head; it is claiming its training recipe reaches comparable quality for a fraction of the compute cost. According to a synthesis of the announcement, the 4.2x-to-12.3x figure comes from fitting scaling-law curves to estimate how much training compute the MAI-Base-2 recipe would have needed to match DeepSeek-V4-Pro-Base's quality, then dividing by what DeepSeek actually spent — and Microsoft says it spent that saving on training a larger model rather than stopping at a smaller one. The company also claims a 6.3x-to-12.7x efficiency improvement over its own MAI-Base-1.

On the actual quality measure Microsoft used — bits-per-byte on held-out text evaluations, where lower is better — the company says MAI-Base-2 beats DeepSeek-V4-Pro-Base as well as Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. But the company states its own caveat openly: only DeepSeek released a true base checkpoint, so the others are post-trained models being compared against a base model, which is an imperfect comparison. And as the coverage notes, no outside research group had replicated any of the efficiency numbers as of October 16 — the estimates are Microsoft's own, derived from its own evaluation sets, with code drawn from its own codebase and contamination-limited STEM, math and QA sets commissioned from vendors. Whether those sets represent the tasks frontier models face in deployment remains unknown.

Suleyman's town-hall shot at OpenAI, and what comes next

The internal message traveled faster than the technical one. On Thursday, Suleyman told employees in a town hall that the company's new pre-trained base model is stronger than the model currently powering OpenAI's GPT-5.6 Sol, and that it was built with roughly a quarter of the compute OpenAI used — according to reporting on his remarks, citing Business Insider. He reportedly said Microsoft privately benchmarked MAI-Base-2 directly against OpenAI's comparable system, internally nicknamed Spud, using technical information Microsoft accesses confidentially. The timing is the real story: Microsoft remains OpenAI's biggest financial backer, with a stake likely worth more than $100 billion, while simultaneously engineering its own way out of dependence on OpenAI's expensive frontier models for Azure and Copilot products.

The arc behind MAI-Base-2 is one of self-sufficiency. Until late 2025, Microsoft was contractually barred from training frontier models that would compete with OpenAI's GPT series; the contract amendment that lifted that restriction was followed by the announcement of seven in-house MAI models at Microsoft Build 2026 in May. MAI-Base-1, which underpins the company's first frontier reasoning model MAI-Thinking-1, was the previous generation of that effort.

What has not been disclosed is arguably as important as what has. MAI-Base-2 is still in post-training — the instruction tuning, reinforcement learning and safety work that turns a base model into a deployable product. Microsoft has announced no release date, no API price and no list of Foundry partners that will receive weight licenses. And the number the agent ecosystem is watching for — performance on real agentic benchmarks, the multi-step, tool-using workflows that enterprise buyers actually run — has not been published. A base model with excellent bits-per-byte does not automatically become an excellent agent; that translation happens in post-training, and the results of that work have not shipped. Still, the direction is clear: Microsoft now trains at a scale it could not legally attempt a year ago, and its next move will be measured not in FLOPs saved but in agents built on what it learned spending them — whether those agents end up transacting over private payment rails built for AI agents or running inside companies rethinking what frontier labs owe the public, like Anthropic's 5% public-benefit share pledge.