Microsoft's MAI-Base-2 Has 2.8 Trillion Parameters. What Microsoft Claims, and What a Base Model Does Not Tell You

On October 15, Microsoft AI published a post about MAI-Base-2, a new base model it says has 80 billion active and 2.8 trillion total parameters.

It is a technical post about pre-training, not a product launch. That distinction matters for builders. This article covers what Microsoft states, what it says about how the model compares, and what you cannot conclude from it. Everything comes from Microsoft's post, so the comparisons are the vendor's own.

What Microsoft says

According to the post:

• MAI-Base-2 is a mixture-of-experts foundation model, with 80B active and 2.8T total parameters.
• It was trained on 16,000 GB200 GPUs over 38 days in August and September 2026.
• Average model FLOPs utilization (MFU) was 21%, and goodput, the share of time training made progress, was 89%. Microsoft says MFU is normalized by 2.5 PFLOPS per GB200 in BF16 and excludes rematerialization and optimizer FLOPs.
• It follows MAI-Base-1 and the June model MAI-Thinking-1, which was post-trained on top of MAI-Base-1.
• MAI-Base-2 is currently undergoing post-training, and Microsoft says it will share more results soon.

The comparison claims

Microsoft says MAI-Base-2 outperforms every open-weight model it tested on its held-out evaluations, and achieves lower bits-per-byte (BPB, lower is better) than the others on those tests. It says the model is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which it calls the strongest open base model.

Some details about how it compared, from the post:

• It compared against checkpoints available before October 2026: DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL.
• For the last three, only post-trained checkpoints exist, so Microsoft notes post-training may affect the results.
• Held-out code evaluation uses Microsoft's own codebase. STEM, math and Q&A evaluations were commissioned from vendors to reduce contamination risk, and all sets were decontaminated against its training data.
• It says DeepSeek-V4-Pro-Base was trained on significantly fewer FLOPs than MAI-Base-2, which is why it estimates efficiency separately from raw results.

These are measurements Microsoft designed and ran. I have not seen independent replication, and I did not review the figures in the post.

The training ideas it highlights

Microsoft names three initiatives: scaling science, model-infrastructure co-design (improved MoE and Muon implementations, and activation gradient clipping), and scaling data, including vision data.

One technique it describes is a scaling ladder: a family of models of different sizes with a fixed recipe, where the layer count determines the architecture. To handle repeated data at small scale, it uses what it calls source scaling, which restricts the unique samples so each dataset repeats the same number of times regardless of run length. Microsoft says this helped results transfer across scales and was key to training a 2.8T model reliably.

Licensing and data claims

The post says Microsoft's pre-training is built on organic data and does not involve distillation from other models. It says this provenance will let Microsoft license model weights to trusted partners through Foundry. It does not say when, or on what terms.

What you cannot conclude

• It is not available to use. The post describes a base model still in post-training. I found no release date or pricing in it.
• Base model results are not chat or agent results. BPB measures how well a model predicts held-out text. It does not tell you how well it follows instructions or runs tools.
• Efficiency claims are relative to a chosen reference. The 4 to 12 times figure compares against one open model, using Microsoft's fitted scaling laws.
• Size is not quality. I am not drawing a ranking from parameter counts.

What I would watch for

These are my own suggestions.

1. A post-trained model release with an API or Foundry listing.

2. Independent benchmark results from third parties.

3. The weight-licensing terms through Foundry, if and when published.

4. Pricing, since a larger model changes the cost picture for hosted use.

Frequently asked questions

What is MAI-Base-2?

A mixture-of-experts base model from Microsoft AI with 80B active and 2.8T total parameters, per its post.

Can I use it today?

The post says it is undergoing post-training. It does not announce availability.

How was it trained?

On 16k GB200 GPUs over 38 days in August and September 2026, with 21% average MFU and 89% goodput, according to Microsoft.

How does it compare to other models?

Microsoft says it beats every open-weight model it tested on held-out evaluations, using BPB. These are its own results.

What is the efficiency claim?

That it is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base.

Will the weights be licensed?

Microsoft says it expects to license weights to trusted partners through Foundry. No terms were given in the post.

Does this tell me how good it is at agents or coding?

No. It reports pre-training results, not post-trained behavior.

Sources

• Microsoft AI, Frontier Pre-training at Microsoft AI: all claims above (October 15, 2026).
• TLDR AI, October 17 issue: discovery pointer only.