For years, when you used Copilot, Microsoft's assistant, it was actually an OpenAI model answering you. Microsoft had invested billions in OpenAI, and had made it the engine of its products.
This week, Microsoft showed that it now builds its own engine. And not a small one.
What was announced
Microsoft's AI lab, led by Mustafa Suleyman, one of the co-founders of DeepMind, announced on Thursday that it had finished pre-training MAI-Base-2. It is a model with 2,800 billion parameters, of which 80 billion work on each word: only part of the model activates depending on the question, like a large hospital where you only consult the specialists you need.
To build it, Microsoft ran 16,000 Nvidia chips for 38 days, in August and September, on a set of 50,000 billion tokens, nearly a fifth of which consisted of images paired with text. Its knowledge stops in March 2026.
It is the second model of this kind for Microsoft. The first, MAI-Base-1, almost three times smaller, serves as the basis for a reasoning model presented in June.
A milestone, not yet a product
Careful not to confuse the two: MAI-Base-2 is not yet an assistant you can use. Pre-training is the phase where the model learns language and the world by reading colossal amounts of text. Then comes post-training, which teaches it to answer usefully, to reason and to behave properly. We explained these stages in our article on what happens after pre-training.
Microsoft gave neither a release date nor a price. Its parameters will not be downloadable: the company plans to offer them under licence to trusted partners, via its Foundry platform.
What Microsoft claims
According to the company, MAI-Base-2 beats all the open models it has tested, on its own evaluations. It also claims that its training method is 4 to 12 times more efficient than that of DeepSeek, the Chinese benchmark for open models.
Two caveats are in order. These comparisons concern open models, not the best closed models from OpenAI, Google or Anthropic. And they were made by Microsoft itself, on its own tests. We will have to wait for the final version, and for independent evaluations, to know where it really stands.
Microsoft insists on one point: its model was trained without "distillation". Distillation consists of training a model by having it imitate the answers of another, more powerful model. It is an efficient and widespread method, but a controversial one: several American labs accuse competitors of having copied their models this way without authorisation. By claiming entirely "organic" data, Microsoft shields itself from that reproach, and can offer its parameters under licence without fear of a claim.
Why Microsoft is doing this
To stop depending on OpenAI. Relations between the two partners grew complicated as OpenAI became an autonomous giant. Having your own frontier model means keeping control of your products, whatever happens.
To keep costs under control. Paying for outside models is expensive, as Microsoft's decision to put its own engineers on an AI diet showed. A homegrown model, once built, changes the equation.
To stay in the race. With Meta having just released Muse Wonder, all the big tech groups want their own frontier model. Microsoft did not want to be the only one depending on someone else.
What we take away
Microsoft was long OpenAI's closest ally. Little by little, it is becoming a competitor.
MAI-Base-2 is still only a foundation, and its real performance remains to be proven. But the message is clear: in AI, no giant wants to depend on another for its engine any more.