Microsoft Finishes MAI-Base-2, a 2.8T-Parameter AI Model

Microsoft AI said on October 15, 2026 that it has finished pre-training MAI-Base-2, a successor it describes as "a larger, sparser model than MAI-Base-1"—a 2.8 trillion-parameter mixture-of-experts (MoE) system with 80 billion active parameters per token. Investing.com framed the announcement as part of Microsoft's push to build its own frontier models and reduce reliance on external providers.

Key facts

What Microsoft announced

In a post titled "Frontier Pre-training at Microsoft AI" and an accompanying technical PDF, Microsoft AI described MAI-Base-2 as the next step in a deliberate frontier pre-training program. The model is still in post-training; Microsoft said it looks forward to sharing more results soon and that preparations for MAI-Base-3 are already underway.

The company framed the release less as a consumer product drop and more as evidence that its internal recipe—data, systems, and MoE architecture—can scale. Secondary coverage from Investing.com on October 15 characterized the move as part of Microsoft's push to build its own frontier models and reduce reliance on external providers. Microsoft's own post did not spell out commercial terms or name licensing partners.

On benchmarks Microsoft controls, the team said MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations, measuring bits-per-byte on held-out code, STEM, math, and Q&A. Comparisons included DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL—with the caveat that only post-trained checkpoints exist for the latter three rivals, so the comparison is base model against partially finished products in some cases.

MAI-Base-2 vs MAI-Base-1

MAI-Base-1 arrived in June 2026 as a 1 trillion-parameter MoE with 35 billion active parameters. MAI-Thinking-1, also from June, was post-trained on that base. MAI-Base-2 roughly triples total scale and more than doubles active parameters while changing how MoE is applied throughout the stack.

Microsoft attributed part of the jump in scale to efficiency gains: the MAI-Base-2 recipe is 6.3–12.7× more pre-training FLOPs-efficient than MAI-Base-1's recipe, and the team reinvested that headroom into training a larger model rather than stopping at a modest size increase.

How it was trained

The pre-training run used 16,000 NVIDIA GB200 GPUs for 38 days in August and September 2026. Microsoft reported an average model FLOPs utilization (MFU) of 21% and goodput of 89% across the job—figures that matter because frontier MoE training is as much an engineering contest as a data contest.

The text-and-vision mixture totaled 50 trillion tokens. Microsoft said the corpus grew about 36% in unique tokens versus MAI-Base-1, with a March 2026 knowledge cutoff. Vision-language data accounted for 18% of the mixture; the VLM portion alone comprised 4.5 trillion unique tokens, covering chart understanding, visual grounding, computer use, medical understanding, sheet-music reading, and related tasks.

On data provenance, Microsoft stressed organic data which does not involve distillation from other models, with every token processed in-house and a proprietary crawler that respects robots.txt. That line is central to how the company plans to commercialize weights later.

Engineering highlights from the post include raising MoE-layer MFU from 24% to 35% via multi-round dropless MoE with fixed memory, switching from AdamW to the Muon optimizer (Muon accounted for 2.1% and 165 ms of average iteration time), and activation gradient clipping for stability at this scale.

The efficiency claim and its caveats

Microsoft's strongest efficiency statement compares pre-training FLOPs against DeepSeek-V4-Pro-Base, which it calls the strongest open base model in its tests. Microsoft estimates its recipe is 4.2–12.3× more FLOPs-efficient than that baseline—a wide band that reflects different accounting choices across the training stack.

Readers should treat benchmark leadership and efficiency multiples as Microsoft's internal claims until independent labs reproduce them. The held-out eval suite is not the same as public leaderboards; the open models named in the post may only exist today as post-trained checkpoints, which skews apples-to-oranges comparisons. None of that negates the training scale—2.8T parameters on 16k GB200s is itself a statement—but it does mean headline "beats every open model" language deserves scrutiny.

Why Microsoft wants its own frontier model

Microsoft has long paired cloud AI services with partner models, but MAI-Base-2 is explicitly an in-house foundation layer. The company said clear data provenance will enable us to license our model weights to trusted partners through Foundry—aligning model weights with the same Foundry branding used for hosted models and tools, without announcing pricing or partner names yet.

That licensing angle sits alongside Microsoft's hardware push. The same week's news cycle includes Microsoft's Maia 300 AI chip roadmap toward multi-gigawatt deployments in 2027, suggesting a vertically integrated story: custom silicon, proprietary data, and frontier-scale MoE checkpoints that Microsoft controls end to end.

On the open-weights side, rivals are shipping massive MoE checkpoints for local and cloud inference—Meta's Muse Spark Open is one high-profile example from October 15. MAI-Base-2 is not an open release; it is Microsoft's bet that efficiency and data cleanliness will matter as much as parameter count when enterprises choose whose weights they can legally fine-tune and deploy.

What happens next

MAI-Base-2 remains in post-training. Microsoft promised more results soon but did not set a release date, API tier, or weight-access program. Foundry licensing for trusted partners is a stated direction, not a live catalog entry.

Internally, MAI-Base-3 planning is already underway—typical for labs that pipeline data collection, architecture experiments, and cluster reservations months ahead of a run. For developers watching Microsoft's stack, the near-term signals are post-training benchmarks, any Foundry partner announcements, and whether the next Maia generation shifts more training off third-party GPUs.

Bottom line

Microsoft AI finished pre-training MAI-Base-2, a 2.8 trillion-parameter MoE with 80 billion active parameters, trained on 50 trillion tokens across 16,000 GB200 GPUs in 38 days. The company claims large efficiency gains over its previous base recipe and over leading open base models, plus leading results on its own held-out evals—with important caveats about independent verification and rival checkpoint maturity. Weights are not public yet; post-training and a path to Foundry licensing come next.

Sources

• Microsoft AI: Frontier Pre-training at Microsoft AI — Oct. 15, 2026
• Microsoft AI: Frontier Pre-training at Microsoft AI (PDF)
• Investing.com: Microsoft AI unveils 2.8 trillion-parameter model, claims efficiency gains — Oct. 15, 2026