Microsoft's MAI-Base-2: 2.8T Parameters, 80B Active and a 4-12x Efficiency Claim

Microsoft has finished pre-training MAI-Base-2, the larger successor to its in-house MAI-Base-1. It is a mixture-of-experts model with 2.8 trillion parameters in total, but only 80 billion of them work on any given token. By our calculation, that is fewer than 3% of the model's weights active at once.

The headline claim comes from Microsoft itself: it says the model is "4-12x more pre-training FLOPs-efficient" than DeepSeek-V4-Pro-Base, which it calls the strongest open base model. You can't use MAI-Base-2 yet. It is still being post-trained, and Microsoft says it will go into its own products first.

Based on Microsoft AI's "Frontier Pre-training at Microsoft AI" post, Business Insider, Investing.com and Bloomberg reporting, checked October 16, 2026. We have not tested the model. All performance and efficiency figures are Microsoft's, and none have been independently verified.

Base model vs. chatbot: what Microsoft actually built

A base model is the raw result of pre-training. In that first stage the model reads a huge amount of text, images and code and learns to predict what comes next. It knows a lot, but it isn't an assistant yet.

The Copilot-style assistant you chat with comes from post-training. That is a second stage where the base model is tuned to follow instructions, reason step by step, decline harmful requests and use tools. In June Microsoft announced MAI-Thinking-1, which it describes as "a general-purpose foundation model post-trained on top of MAI-Base-1."

Business Insider explains why the first stage matters. Pre-training "takes massive amounts of expensive computing power, so Microsoft could save money if it's worked out a way to do this more efficiently."

The specs, according to Microsoft

• Size: 2.8 trillion total parameters, 80 billion active, in a mixture-of-experts design. Microsoft calls it "a larger, sparser model than MAI-Base-1, trained on more data."
• Training data: a 50-trillion-token mix. Vision-language data made up 18% of it, and Microsoft says the model was pre-trained natively with vision rather than having vision bolted on later. Its web data runs to March 2026.
• Hardware: 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026. Microsoft reports 21% average model FLOPs utilization and 89% "goodput," the share of time training was actually making progress.
• Data source: Microsoft says its pre-training is built on "organic data which does not involve distillation from other models."

Microsoft doesn't publish a total compute figure, so we estimated one. Multiplying 16,000 GPUs by the 2.5 petaFLOPS peak Microsoft quotes for the GB200, then by 21% utilization, 89% goodput and 38 days, gives roughly 2.5e25 FLOPs. A common rule of thumb (six times active parameters times tokens) gives 2.4e25. That is about 14.6 million GPU-hours. These figures are our arithmetic, not Microsoft's.

What "4x-12x" is actually measured against

This claim is easy to misread. It doesn't mean MAI-Base-2 is 4 to 12 times smarter, cheaper to run or faster than DeepSeek. It is a claim about how much training compute Microsoft's recipe needs to reach the same quality as one specific open model.

Here is how Microsoft says it worked it out. It trained a "scaling ladder," a family of models of different sizes using one fixed recipe. On four held-out evaluations it plotted bits-per-byte, a measure of how well a model predicts unseen text, against training compute. It then estimated the compute a ladder model would need to match DeepSeek-V4-Pro-Base's score and compared that with DeepSeek's training compute. The detailed figure is "4.2–12.3x more FLOPs-efficient," which the summary rounds to 4-12x.

Microsoft says its held-out code test uses its own codebase, while the STEM, math and Q&A tests were commissioned from outside vendors and checked against its training data for contamination. It also claims "MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations," and a 6.3-12.7x efficiency gain over MAI-Base-1, which was released in June.

The OpenAI angle, and what it means for Copilot

Microsoft's post doesn't mention OpenAI. Chief executive Mustafa Suleyman did, inside the company. At a town hall on Thursday, Business Insider reported, he told employees: "We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol. And in fact, we used one-quarter of the compute that OpenAI used to train their model."

According to Business Insider, Suleyman said Microsoft privately compared its model with OpenAI models, including one nicknamed "Spud" that powers GPT-5.6, using technical information Microsoft holds about them. None of those results are public. Microsoft declined to comment to the outlet, and OpenAI did not respond.

The context explains why this matters. Business Insider notes that Microsoft owns an OpenAI stake "likely worth more than $100 billion," sells OpenAI models to Azure customers and has used them in products such as Copilot, but that the relationship "has frayed in the past year or so." In July, Bloomberg reported that Microsoft had started routing some Excel and Outlook prompts to its own MAI models instead of OpenAI and Anthropic models. That was still a small share of overall AI usage, and cost was the main driver.

For Copilot users, very little changes for now. Microsoft says owning the full stack lets it tune models and infrastructure "for the best performance at the lowest cost, first in our own products and then for Foundry customers." It hasn't named products, given a date or said whether Copilot will use a model built on MAI-Base-2.

Developers won't get open weights. Microsoft says it plans to "license our model weights to trusted partners through Foundry," and links that to the no-distillation training data. Investing.com reports that Microsoft has not disclosed commercial availability dates or pricing. No public API has been announced.

If you manage Copilot for a business, watch Microsoft's model-selection notes and the Foundry catalogue for any switch.

The catch: a strong base model is only half the job

Bits-per-byte measures how well a model predicts text. It isn't a measure of how well a chatbot answers your question, writes working code or follows instructions. Microsoft publishes no downstream task scores for MAI-Base-2, and the useful version of the model doesn't exist yet.

Microsoft flags two limits itself. It says direct model comparisons "conflate the quality of the training recipe with differences in model size." Three of the rivals it tested were only available as post-trained checkpoints, which "may affect these evaluations." The code test also uses Microsoft's own codebase, which outsiders can't reproduce. The OpenAI comparison rests on unpublished private data.

Frequently asked questions

Can I use MAI-Base-2 today?

No. Microsoft says it is still in post-training. It plans to use it in its own products first, then license the weights to trusted partners through Foundry. There's no public release, API or price yet.

Is MAI-Base-2 open source?

No. Microsoft compares it with open-weight models but hasn't said it will release the weights publicly. It plans to license them only to selected partners.

Was it trained on Microsoft's own chips?

Microsoft says it trained the model on 16,000 Nvidia GB200 GPUs. The post doesn't mention Microsoft's in-house Maia chips.

Microsoft says preparations for MAI-Base-3 are already under way. The real test will be the post-trained model, and whether outside testers can confirm Microsoft's efficiency claims once it ships.