Microsoft AI announced completed pre-training of MAI-Base-2 on 15 October 2026: a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters. The model is still undergoing post-training. This is a development report, rather than a release of a finished public assistant or freely downloadable weights.
The company compares models using bits-per-byte on held-out data and acknowledges that some comparison checkpoints have undergone post-training. Its claimed improvement in compute efficiency concerns pre-training. It does not promise the same reduction in cost or time for every end-user task.
Why the development stage matters
Pre-training creates a foundation on which assistant behaviour can subsequently be built. For a reader, the useful distinction is between predicting data and reliably carrying out a particular request. Strong performance on the first task can be a meaningful research signal without constituting a completed evaluation of the second.
Consider a hypothetical internal assistant that must locate the correct document and answer with a verifiable source. A model might be strong at language prediction while the application still needs to select the right document, preserve access permissions and distinguish an unknown answer from a plausible addition. Increasing the total parameter count does not remove those requirements from the surrounding system.
Our editorial interpretation is to consider the foundation, post-training and application separately. When user-facing evaluations appear, readers should be able to see which layers participated. A comparison with another product can involve different tools, reasoning time and information access. Without that frame, a table can easily support a broader conclusion in the reader's mind than the measurement actually permits.
What the chosen metric tells us
Bits-per-byte evaluates content prediction relative to bytes, avoiding a direct comparison between incompatible token units. It helps pose a more explicit question about language modelling. It does not turn the evaluation into a measurement of completed orders, correctly edited files or user satisfaction. The operational task remains different from the statistical quantity being evaluated.
We would read a lower value as a result within the specific assessment, rather than a universal ranking. The choice of test data and possible overlap with training remain important. When one checkpoint is a base model and another has been adapted to follow instructions, a comparison should carry that qualification rather than leaving readers to infer it from product names. That distinction matters even when all reported numbers are accurate.
Training efficiency and product efficiency
Training a foundation with less computation for a given quality is a different objective from serving real traffic cheaply. In a finished product, context size, concurrency, response time and retries after failures also affect cost. We do not derive those operational quantities from the pre-training account. For a team choosing a provider, the next relevant evidence will be an actually available model and its behaviour on the team's own task.
The hypothetical document assistant, for example, needs a final answer that is grounded in accessible material. An evaluation would need to record whether that happened, how much correction was required and what resources were used. Those are questions for a later deployment, not results we assign to MAI-Base-2 from a development announcement.
SIExpo has not reproduced the training or evaluations. The significant event is progress on a large in-house model foundation, with the boundary between a completed development stage and a future product kept explicit.
Editorially checked: 2026-10-16T16:26:59.575073+00:00 (UTC). Cover: conceptual AI illustration.
