Microsoft has finished pre-training MAI-Base-2, a model with 2.8 trillion total parameters. In an October 15 research post, the company says it beats the open-weight models it tested and was trained with a more efficient recipe than DeepSeek's comparable base model. It isn't a usable assistant yet. Microsoft says it's "currently undergoing post-training," the work that turns a base model into one.
If you use Copilot or buy AI through Microsoft, the real question is how much of that service will eventually run on models Microsoft trains itself. The company says it wants its own pipeline in its products, and plans to license the weights to trusted partners through Foundry.
What "4 to 12 times" actually measures
That figure has been repeated a lot, and it's about training. It's Microsoft's estimate of how much less computation it needed to reach particular evaluation scores. It doesn't mean MAI-Base-2 answers questions twelve times faster, and it doesn't mean your API bill will drop by that much.
Microsoft's technical report shows the working. It fits scaling curves to models trained with its recipe, then estimates how much computation it would take to match DeepSeek-V4-Pro-Base on four held-out evaluations. Microsoft credits the gain to its data and modeling choices, and says it spent the savings on training a bigger model.
That's the catch for customers. A more efficient recipe can either cut the cost of a given result or buy a better result for the same money. Microsoft chose the better result, so there's no price cut you can work out from this.
A good base score isn't a good assistant
The evaluations measure how well the model predicts held-out text, including code, maths and question-and-answer material. They use a metric called bits per byte: lower means the model predicted the text more accurately. Researchers like it because it compares models that split text into tokens differently.
What it doesn't test is whether an assistant built on the model will get your spreadsheet right, ask before sending a message, or recover when a tool call breaks. That depends on post-training and on the software wrapped around the model.
The report also flags a weakness in its own comparison. Some rival models had already been post-trained, because their base versions weren't available. So the chart is evidence about the evaluations Microsoft chose. It can't tell you which finished assistant will work best for your team.
Microsoft wants control of its own supply
"Owning our model pipeline end-to-end allows Microsoft to make a distinctive offering to our customers," the post says. Owning a model means Microsoft can change how it's trained, tune it for a specific product and coordinate how it runs on its own infrastructure. It also says its pre-training "does not involve distillation from other models," a point it connects to future licensing.
For a business customer, the upside is another source of models inside Microsoft services you already use. That only helps if it comes with clear information on what the model can do, how it handles your data and what support you get. A new model name on its own changes none of that.
It also connects to Satya Nadella's call for controls outside the model. Training a better predictor and building an agent you can trust are related jobs. The second still needs permissions that hold, logs you can inspect and a way to stop it mid-task.
So don't change a production workflow yet. Wait for the post-trained model's results and access terms, then test it on the jobs your current model already does. Today, MAI-Base-2 is a research result. It becomes a buying decision the day you can try it.
Cover: AI-generated editorial illustration by TMRW using Microsoft's official logo. The processor board is illustrative and does not depict an MAI hardware product.
