Microsoft finishes training MAI-Base-2, a 2.8-trillion-parameter AI model

MAI-Base-2 is a foundation model: the raw, pre-trained version of an AI before anyone teaches it manners. Microsoft calls it state of the art and says it is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active ones.

Parameters are the adjustable numbers inside a model that hold what it learned. More parameters can mean more stored knowledge. Microsoft says it has a knowledge cutoff of March 2026.

Here are the headline numbers in one place, from Microsoft AI's technical post:

• Total parameters: 2.8 trillion
• Active parameters per token: 80 billion
• Hardware: 16,000 Nvidia GB200 GPUs
• Training time: 38 days, in August and September 2026
• Training data: a mix of 50 trillion tokens
• Knowledge cutoff: March 2026

Mustafa Suleyman, who leads Microsoft AI, shared the news on X, rounding it up to "3T-parameter weight class" models "built entirely in-house," according to Pasquale Pillitteri's write-up. The technical post gives the exact figure: 2.8 trillion.

A mixture-of-experts model splits its network into many smaller specialist parts, called experts. For each word it reads or writes, a router picks only a few of them to do the work.

Think of a huge hospital. It employs thousands of doctors, but you only see the two or three who fit your problem. The hospital holds a lot of knowledge, yet your visit doesn't cost the time of every doctor in the building.

That's why the two numbers matter. The 2.8 trillion total says how much the model can hold. The 80 billion active says how much computing each word costs. Only a small slice of MAI-Base-2 is awake at any moment, and that's the whole trick.

MAI-Base-2 replaces MAI-Base-1, the base model behind MAI-Thinking-1, the reasoning model Microsoft announced in June. TokenPost reports the jump in size:

The new model is bigger and sparser, which means a smaller share of it is active at once. It also used nearly twice as many GPUs, TokenPost notes.

One more change: MAI-Base-2 learned from images from day one. Microsoft says vision-language data made up 18% of the training mix. TokenPost puts that at about 9 trillion tokens in the mix, drawn from 4.5 trillion unique vision tokens. Microsoft also says its pool of unique training data grew by about 36% since MAI-Base-1.

Microsoft makes two claims. First, MAI-Base-2 outperforms every open-weight model the company tested on its held-out evaluations. Second, it is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which Microsoft calls the strongest open base model.

FLOPs are a count of the math a computer does. More efficient means Microsoft got the same quality for less computing, and it says it spent that saving on training a bigger model.

By the numbers: the exact range is 4.2 to 12.3 times, measured across four held-out evaluations, according to the post. Those tests cover code, question answering, math and STEM. Microsoft scores them in "bits per byte," where lower is better. It says it commissioned some test sets from outside vendors and checked them against its training data so the model hadn't seen the answers.

The catch: these are Microsoft's own measurements. TokenPost points out that some rivals in the chart, like Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, only had post-trained versions available, which may affect the comparison. Nobody outside Microsoft can test MAI-Base-2 yet.

Microsoft credits three efforts in its post.

The post also lists the lab's guiding principles. Three stand out: "intelligence should be learned from first principles," "simple methods scale best," and "scientific rigor over shortcuts."

• Scaling science: "scaling ladders" that predict how a big model will behave from smaller test runs, adjusted for data the model sees more than once.
• Model and hardware designed together: experts in every layer instead of every other layer, a faster version of an optimizer called Muon, and a trick called activation gradient clipping that keeps training stable.
• More data: bigger data pipelines and image data built in from the start.

Some of the speedups are big. Microsoft says fixing inefficiencies in the expert layers raised how much of the chips' power it actually used from 24% to 35% in that part of the work. Across the full run, it reports 21% average utilization and 89% "goodput," the share of time the training actually moved forward.

The Muon optimizer, the piece that decides how to adjust the model after each step, took 2.1% of each training step, or about 165 milliseconds, Microsoft says. That's 165 thousandths of a second, repeated across 38 days. Somewhere a GPU is very tired.

Muon works by reshaping the update for each block of the model's weights with a short chain of matrix math, Microsoft explains. The company estimates the step could, in theory, run in about 55 milliseconds, so there is still room to speed it up. One fix it describes is spreading the optimizer's memory across more chips.

The clipping trick matters more as models grow. Microsoft says it keeps the training signal in a much tighter range than it did for MAI-Base-1, and that the benefit increases with scale.

Microsoft says its pre-training is built on organic data with no distillation from other models. Distillation means training a model on another AI's answers. Microsoft says it processes all training tokens in-house.

The company says it collects web data with its own crawler that respects robots.txt, the file websites use to say "please don't scrape this," plus related page-level controls. It also says it leaves out sources that break its Responsible AI policies or appear on the US Trade Representative's Notorious Markets list.

In real life If you run a website and block AI crawlers in your robots.txt file, Microsoft says its crawler should skip your pages when gathering training data.

What's next: MAI-Base-2 is now in post-training. That's the stage that turns a raw model into one that follows instructions, chats and refuses harmful requests. Microsoft says it will share more results soon.

After that, Microsoft plans to license the model's weights to trusted partners through Foundry, its platform for selling AI models to businesses. TokenPost notes that no partners or release date have been announced.

And the next one is already in the works. Microsoft says preparations for MAI-Base-3 are underway. We last covered the lab when it launched real-time transcription and voice models, and other giant models like Mistral Large 4 show how crowded the trillion-parameter club is getting.

• Nothing changes today: MAI-Base-2 isn't in Copilot or any app you use. It's still being trained to talk.
• Watch Microsoft products: MAI-Thinking-1 grew out of MAI-Base-1, so models built on MAI-Base-2 could reach Microsoft's tools and Foundry later.
• Businesses come first: the plan is to license the weights to trusted partners, not to release them for anyone to download.
• Treat the benchmarks with care: the comparisons are Microsoft's own, and outside testing hasn't happened yet.