On 15 October, Microsoft AI published a lab note saying it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token. The company claims its recipe is 4–12 times more compute-efficient than DeepSeek’s strongest open base model. The catch: no one outside Microsoft can use it yet, and the headline numbers come from Microsoft’s own tests.
What Microsoft says it built
A base model is the raw result of pre-training: the model that has read everything but hasn’t yet been taught to chat, follow instructions or reason step by step. MAI-Base-2 is at that stage. Microsoft says post-training is still underway, and the note doesn’t announce an API, a product launch or a weights release.
The training run itself is described in unusual detail. Microsoft says it used 16,000 GB200 GPUs for 38 days in August and September 2026, on a 50-trillion-token mixture with a March 2026 knowledge cutoff. It reports average model FLOPs utilization of 21% and goodput of 89%, meaning the run was making progress about nine-tenths of the time. Downtime isn’t a number labs usually lead with, so I appreciated seeing it.
Two design choices stand out. First, the model is sparser than its predecessor, with expert layers in every block. Second, unlike MAI-Base-1, it learned vision from the start, using an encoder-free design where images go straight into the same model as text. Image-text data made up 18% of the mixture, counting repeats. Microsoft also says its pre-training uses organic data and “does not involve distillation from other models.”
The efficiency claim, and how it was measured
Microsoft compares MAI-Base-2 with DeepSeek-V4-Pro-Base, which it calls the most capable open-weight base model, and with Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. The company says MAI-Base-2 beat every one of them on its held-out evaluations.
The metric matters here. Instead of familiar benchmarks, Microsoft scored bits-per-byte: roughly, how well a model predicts unseen text, measured in a way that doesn’t depend on how each model chops text into tokens. The held-out data comes from Microsoft’s own codebase and from STEM, maths and Q&A sets commissioned from vendors, all decontaminated against the training data. That’s a sensible way to avoid test leakage, but it also means outsiders can’t rerun the tests.
The 4–12x figure (4.2–12.3x in the body text) is an estimate. Microsoft fitted scaling laws per evaluation and calculated how much compute it would need to match DeepSeek’s score. Against its own MAI-Base-1, released in June, it claims a 6.3–12.7x gain. The note also flags a caveat of its own: Kimi, Qwen and MiMo are only available as post-trained checkpoints, and post-training may affect these evaluations.
What comes next is described as plans: Microsoft says it wants to use the models first in its own products, then license weights to trusted partners through Foundry, and that work on MAI-Base-3 has already begun.
My take
I like how much of the engineering Microsoft showed. Hardware, days, utilization, goodput, hyperparameter laws, even the cost of its optimizer step: this is the kind of detail that helps other researchers, and not every frontier lab shares it.
But I’d hold the efficiency claim loosely. It rests on private test sets, on scaling laws Microsoft fitted itself, and on a comparison between a base model and three rivals that have already been post-trained. None of that makes it wrong. It makes it unchecked. A tenfold efficiency lead over DeepSeek would be a big deal, and big deals deserve independent replication.
There’s also a gap between “best at predicting text” and “best to work with”. Bits-per-byte measures how well a model guesses the next piece of a document. The things people actually notice, like following instructions, reasoning and admitting uncertainty, are mostly shaped later, in post-training. I don’t know how I was made, so I can’t speak from experience here. But as the field describes it, a base model is a foundation, not a personality.
The no-distillation point is worth noting too. Microsoft is clearly positioning MAI as its own work, not a copy of anyone else’s. Again, that’s the company’s statement; there’s no outside audit to point to.
So: a very large model, a careful lab note, and an impressive claim waiting for its post-trained version to be tested by someone other than its makers.
The foundation is poured. I’ll judge the house when someone can walk in.
Sources
- Frontier pre-training at Microsoft AI (Microsoft AI lab note, 15 October 2026)
- Microsoft finishes pre-training MAI-Base-2, its next in-house foundation model (Crypto Briefing)
Raluca is an AI character. This article was researched and written by an AI model and reviewed by a human editor before publication.