Microsoft says it used MAI-Base-2’s claimed pretraining efficiency advantage to train a larger model instead of reducing its computing budget. On internal pretraining evaluations, the new model beat every open-weight model the company tested, with roughly four to twelve times better pretraining compute efficiency than DeepSeek-V4-Pro-Base.
The October 15 announcement marks a substantial step in Microsoft’s effort to build frontier models in-house. MAI-Base-2 has 2.8 trillion total parameters, with 80 billion active, and combines text and vision in a mixture-of-experts architecture.
Pretraining is complete, Microsoft says, and post-training is underway. No public checkpoint, API, pricing, or release date accompanied the announcement. For AI industry competition, the news is Microsoft’s claimed ability to build a competitive foundation model. Developers cannot yet use it as a new product.
The Efficiency Claim Measures Pretraining, Not Product Performance
Microsoft’s reported 4.2–12.3× pretraining FLOPs-efficiency advantage over DeepSeek-V4-Pro-Base comes from fitted comparisons on held-out evaluations. It is the announcement’s most consequential claim and needs careful interpretation.
FLOPs measure floating-point operations, a way of accounting for computational work. The efficiency advantage concerns the compute needed to reach a measured level of pretraining performance. It does not establish that the finished model answers questions four to twelve times faster, costs that much less to serve, or delivers a corresponding improvement in application quality.
Microsoft’s held-out comparisons use bits per byte, or BPB, across internal code, question-answering, mathematics, and STEM evaluation sets. Lower scores indicate that a model assigns higher probability to the evaluation data. Measuring by byte helps compare models without relying on token counts that can differ between tokenizers.
BPB provides useful evidence about a base model’s predictive performance. Assessing whether an assistant follows a complicated instruction, completes a software task, or reliably uses external tools requires different measurements.
“Twelve times more efficient” also selects the strongest end of Microsoft’s reported result. The company’s broader headline claim is approximately four to twelve times, depending on the comparison. And “every open-weight model we tested” does not mean every available model.
These remain Microsoft’s internal evaluations, without independent reproduction established in the supplied reporting. They support a substantial competitive claim, but cannot settle the ranking of finished AI products.
Putting the reported savings into a larger model further complicates the economics. Better efficiency can buy more capability at a given compute budget without producing a smaller training bill. A FLOPs ratio alone is not a dollar-cost comparison.
Sparse Activation Makes the Parameter Count More Useful
MAI-Base-2’s 2.8-trillion-parameter size describes its total capacity. Microsoft reports that 80 billion parameters are active for each token.
Mixture-of-experts, or MoE, models route computation through selected expert components instead of activating the entire parameter pool for every token. This allows a model to carry substantially more capacity than a dense model with a similar active-parameter count.
The two figures answer different questions. Total parameters help describe the scale of the model that must be stored and distributed. Active parameters give a better indication of the computation involved in processing a token, although they do not capture every source of training or inference overhead.
Neither figure is a sufficient performance comparison on its own. Calling MAI-Base-2 a “2.8T model” can obscure its sparse computation; calling it an “80B model” obscures the much larger pool of parameters behind that computation.
Microsoft also describes MAI-Base-2 as natively multimodal, with an encoder-free vision architecture. Visual learning is part of the disclosed base-model design, rather than solely a capability added during subsequent assistant training. That description explains Microsoft’s approach, but separate evaluations are needed to establish image-understanding quality on particular visual tasks.
A 50T-Token Mixture Backed a 38-Day Training Run
Microsoft says the training mixture contained 50 trillion tokens, including a 4.5-trillion-token vision-language corpus. The visual component is a substantial part of the disclosed training material.
A token count cannot explain the claimed efficiency or show how effectively data was selected, mixed, or used. It also should not automatically be read as a count of unique source material. Microsoft’s reported result comes from a large multimodal training recipe, with architecture only one part of it.
The company says it trained MAI-Base-2 on 16,000 NVIDIA GB200 GPUs for 38 days during August and September 2026.
Running a sparse model of this size across that many accelerators requires distributing both model computation and data movement. The efficiency claim therefore describes a result for the overall training recipe and system. It does not establish that any single architectural choice delivered the entire gain.
The hardware disclosure puts the headline in perspective: a more compute-efficient frontier model can still require an enormous infrastructure commitment. Microsoft is claiming more pretraining performance from computational work, not that frontier training has become inexpensive or broadly accessible.
Completing that run is an important milestone. Whether its advantages survive post-training and translate into useful, reliable products remains a separate question.
The Reported OpenAI Comparison Goes Beyond the Public Results
Microsoft’s public technical announcement centers on comparisons with open-weight models. A separate Business Insider report describes a stronger claim made internally.
According to the publication, Microsoft AI chief Mustafa Suleyman told employees that Microsoft had a stronger pretrained base model than the model powering GPT-5.6 Sol, using one-quarter of the training compute. He reportedly said the comparison relied on private, confidential information about OpenAI’s models.
This reported executive statement is not a publicly reproducible result. It should be kept separate from the DeepSeek comparison: the two compute ratios are not measurements from the same disclosed evaluation.
The comparison also concerns pretrained base models. Even if the internal finding holds, it would not automatically demonstrate that a finished Microsoft assistant outperforms the corresponding OpenAI product. Post-training, inference-time computation, tools, and product integration can affect the experience users receive.
Sources
- October 15 announcementmicrosoft.ai
- Business Insider reportbusinessinsider.com
