Microsoft MAI-Base-2: a 2.8-trillion-parameter base model, a 4–12x efficiency claim, and no release date

On October 15, 2026, Microsoft AI said it had finished pretraining MAI-Base-2, the base model for its next generation of in-house frontier models. It's a mixture-of-experts (MoE) model with 2.8 trillion total parameters and 80 billion active per token, trained on 16,000 Nvidia GB200 GPUs over 38 days. Microsoft claims it's 4 to 12 times more efficient to pretrain than DeepSeek's strongest open base model. You can't use it yet: MAI-Base-2 is still in post-training, and Microsoft plans to license it only to "trusted partners" through Foundry.

This story is based on Microsoft AI's October 15 technical post, Microsoft's Foundry catalog, DeepSeek's Hugging Face page, and reporting by Business Insider and StreetInsider, checked October 16, 2026.

The short version

• What: a pretrained base model, not a product. It's the foundation Microsoft will post-train into chat, reasoning and coding models.
• Size: 2.8T total parameters, 80B active, with MoE in every layer.
• Training: a 50-trillion-token mixture with a March 2026 knowledge cutoff, on 16,000 GB200 GPUs for 38 days in August and September 2026.
• Claim: 4.2–12.3x more pretraining-FLOPs-efficient than DeepSeek-V4-Pro-Base. Vendor-reported.
• Availability: none yet. Microsoft says it will license the weights to trusted partners through Foundry, with no date. It isn't an open-weight model.

What did Microsoft announce?

The post, titled "Frontier Pre-training at Microsoft AI," describes MAI-Base-2 as "a larger, sparser model than MAI-Base-1, trained on more data." Its predecessor, MAI-Base-1, is the base under MAI-Thinking-1, the reasoning model Microsoft introduced in June. Microsoft's Foundry catalog lists MAI-Thinking-1 as a 1-trillion-parameter MoE with 35 billion active parameters and a 256K context, in preview.

Key details from Microsoft's post:

Last verified: October 16, 2026 (Microsoft AI post of October 15)

Microsoft also stresses where its data came from. The post says pretraining "is built on organic data which does not involve distillation from other models," using a proprietary crawler that respects robots.txt. That's notable because distillation, training a model on another model's outputs, has become a flashpoint between labs. It also contrasts with Microsoft's own Microsoft-Decision-1, which is post-trained from Alibaba's Qwen3.5-9B.

How credible is the efficiency claim?

Microsoft claims MAI-Base-2 is "4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base," with a detailed range of 4.2–12.3x. In plain terms: Microsoft says it reaches DeepSeek's level of performance with a fraction of the training compute. It also says MAI-Base-2 "outperforms every open-weight model we tested on our held-out evaluations," naming Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL.

Treat these as vendor-reported, with real limits:

• Microsoft ran the tests. It scored held-out text and its own code itself, and commissioned STEM, math and Q&A sets from outside vendors. Evaluations were decontaminated against its training data, Microsoft says, but they aren't public.
• An estimate, not a head-to-head run. Microsoft says the 4.2–12.3x range comes from fitted scaling laws that estimate how much compute DeepSeek-V4-Pro-Base would need to match MAI-Base-2.
• No absolute scores in the text. The post's comparisons sit in charts; it doesn't list benchmark numbers that others can check. Microsoft says it will share more results "soon."
• Apples and oranges. Microsoft notes that only post-trained checkpoints exist for Kimi K3, Qwen3.8 and MiMo, so it compared its base model against finished models, and post-training may affect results. DeepSeek publishes a base model, so that comparison is more direct.

For context, DeepSeek's Hugging Face page lists DeepSeek-V4-Pro at 1.6 trillion total and 49 billion active parameters, pretrained on more than 32 trillion tokens, under the MIT license.

What did Suleyman reportedly say about OpenAI?

Business Insider's Ashley Stewart reported that Microsoft AI CEO Mustafa Suleyman told an all-hands meeting on October 15 that MAI-Base-2 is stronger than the pretrained model behind OpenAI's GPT-5.6, which BI said is internally called "Spud." According to BI, Suleyman said Microsoft used about a quarter of the compute OpenAI used, and that the comparison relied on technical information Microsoft holds about OpenAI's models. BI also reported that Suleyman said the model used 12 times less compute than DeepSeek's, without naming the DeepSeek model. Microsoft declined to comment to BI, and OpenAI didn't respond.

That claim rests on internal remarks, a private comparison and one outlet's report. It hasn't been published or independently checked, and some secondary write-ups have misspelled the codename. Microsoft's public post doesn't mention OpenAI at all.

Why does it matter?

Microsoft still sells OpenAI's models on Azure and is building agents such as Autopilot on top of Copilot. MAI-Base-2 is the clearest sign yet that it's building a full-size alternative, from data through to its own Maia 300 inference chip, due in 2027. Microsoft says "preparations for MAI-Base-3 are already underway."

For developers, nothing changes today. There's no API, no model card, no price and no post-trained model. The Foundry licensing plan points to a gated model for selected partners, not a public download.

What's confirmed, reported and uncertain

• Confirmed (Microsoft): pretraining is complete; 2.8T/80B MoE; 50T-token mixture; March 2026 cutoff; 16,000 GB200s for 38 days; post-training is underway; plans to license weights to trusted partners through Foundry; MAI-Base-3 preparations.
• Vendor-reported: the 4.2–12.3x efficiency claim against DeepSeek-V4-Pro-Base and the held-out evaluation wins.
• Reported: Suleyman's town hall remarks on OpenAI's "Spud" model and a quarter of the compute (Business Insider, October 15).
• Uncertain: when a post-trained model ships, under what name and price; which partners get weights; and independent results.