On October 15, 2026, Microsoft AI announced that it had completed pretraining of MAI-Base-2, a foundation model with 2.8 trillion total parameters. The model is the base for a next-generation AI system that handles both images and text. According to Microsoft, its training compute efficiency is 4 to 12 times that of DeepSeek-V4-Pro-Base. In practice, however, Microsoft put more compute into MAI-Base-2 than the comparison model used, and post-training to improve instruction following, reasoning, and other capabilities is still under way. What stands out in the technical report is not simply that the model is larger, but that the entire training system was reworked, from how training data is repeated to how GPUs communicate. Below, we look at what Microsoft's claimed efficiency gain means and the technologies used to achieve it.
How do you run a 2.8-trillion-parameter model?
MAI-Base-2 uses a Mixture of Experts (MoE) architecture, which activates only part of the model depending on the input tokens. Total parameters reach 2.8 trillion, but only 80 billion are actually used to process each token. By storing knowledge across a very large number of parameters while limiting which parts take part in each computation, the design keeps the computational load down. This means you cannot compare its computational cost with that of a conventional dense model, which uses all parameters every time, by looking at total parameter count alone.
Microsoft's official technical report summarizes the differences from the previous generation, MAI-Base-1.
| Item | MAI-Base-1 | MAI-Base-2 |
|---|---|---|
| Total parameters | 1 trillion | 2.8 trillion |
| Active parameters | 35 billion | 80 billion |
| Layers | 78 | 102 |
| FFN layout | Alternating dense FFN and MoE | MoE in every layer |
| Training tokens | 30 trillion | 50 trillion |
| Vision pretraining | Text only | Handles images and text |
The FFN (Feed-Forward Network) is the network in each layer that transforms the input features; in an MoE, this processing is distributed across multiple expert networks. MAI-Base-2 selects 7 of 512 expert networks and combines them with a shared component that is always used. Whereas the previous generation alternated dense FFNs and MoE layers, the new model uses MoE for the FFN portion of every layer. This does not mean the attention mechanism or other components were removed; it is a change in how the feature-transforming part of each layer is configured.
Supporting images also involved changes to the model structure, not just additional training data. MAI-Base-2 is designed as a vision-language model without a separate image encoder, and it handles both images and text from the pretraining stage. The vision-language data amounts to 4.5 trillion tokens after deduplication, and including repeated use it accounts for 18% of the total 50 trillion tokens of training. The 50 trillion figure is therefore the total number of tokens processed during training, not all of it distinct data.
The training data also includes material on reading charts and spatial reasoning. However, this describes what the model was trained on, not an evaluation of how much capability it actually acquired. Microsoft had already announced MAI-Thinking-1 in June 2026, so this is not the company's first in-house AI model. MAI-Base-2 is best seen as an expansion of the foundation-model development setup Microsoft has built so far, extended to large-scale pretraining that includes images.
What does "up to 12x training efficiency" actually compare?
The metric Microsoft used to compare model performance is BPB (Bits Per Byte). It expresses prediction loss on text not used in training as bits per byte; a lower value means the model assigns higher probability to that text. It makes it easier to compare models that split text into tokens differently, but it is not a direct measure of how many math problems or programming tasks a model solves correctly.
The evaluation covered four domains: code, Q&A, math, and STEM (science, technology, engineering, and mathematics). For code, Microsoft used MAI's own code as evaluation material; for Q&A, math, and STEM, it says it commissioned external vendors to create the material. Microsoft also removed overlap with the training data, saying this addressed the problem of results being inflated because evaluation items appeared in training data. Still, the evaluation itself is an internal Microsoft evaluation, and the fact that an outside vendor created some of the material does not make it an independent third-party assessment.
Organizing the comparison with DeepSeek-V4-Pro-Base based on Figures 1 and 3 in the October 15 technical report gives the following.
| Domain | DeepSeek-V4-Pro-Base BPB | MAI-Base-2 BPB | Relative BPB reduction | Training FLOPs efficiency ratio (official estimate) |
|---|---|---|---|---|
| Code | 0.2370 | 0.2203 | ~7.0% | 12.3x |
| Q&A | 0.4351 | 0.4162 | ~4.3% | 4.2x |
| STEM | 0.3968 | 0.3653 | ~7.9% | 8.3x |
| Math | 0.3862 | 0.3561 | ~7.8% | 6.6x |
The relative BPB reduction is calculated for each domain as (DeepSeek BPB − MAI BPB) ÷ DeepSeek BPB × 100, rounded to one decimal place. Comparing the finished models, MAI-Base-2's BPB is roughly 4.3% to 7.9% lower than DeepSeek-V4-Pro-Base's. Microsoft's stated efficiency improvement, by contrast, ranges from 4.2x to 12.3x, a far larger figure. This is because the two numbers measure different things, and the BPB reduction cannot be read as an improvement in accuracy.
What the efficiency multiple represents is an estimated ratio of the training compute needed to reach the same BPB. Microsoft trained several models of different sizes using the same training recipe and fitted curves relating training compute to BPB for each domain. From those curves, it estimated the compute required to reach the same BPB as DeepSeek-V4-Pro-Base and compared that with DeepSeek's training compute. The measured values for the finished MAI-Base-2 are also included in the curve fitting, and the comparison point lies within the fitted range.
The important point is to distinguish the efficiency of the training method from the compute actually spent on the finished model. Microsoft itself states that DeepSeek-V4-Pro-Base's actual training compute was considerably smaller than MAI-Base-2's. In other words, it does not mean MAI-Base-2 was trained with one-twelfth of DeepSeek's compute. The company says it redirected the headroom gained from more efficient training methods toward scaling up the model.
The comparison models are also at different training stages. DeepSeek-V4-Pro-Base is a pretrained base model, but others listed for comparison, such as Kimi K3, are post-trained, and Microsoft cautions that this effect must be taken into account. The evaluation description also does not give the number of evaluation items or confidence intervals for the measurements. This report is based on the Lab Notes and technical PDF published by Microsoft, and should be distinguished from a peer-reviewed paper or independent third-party performance verification.
Small-scale experiments reproduce the production data-repetition conditions
Training efficiently on 50 trillion tokens requires first exploring suitable training settings with smaller models. But settings that work well in small experiments are not necessarily optimal for a giant model. To address this, Microsoft matched the production run not only in the number of training tokens relative to model size but also in how many times the same data is reused. Specifically, it prepared models of several sizes with the same number of training tokens per active parameter and unified the data-repetition conditions across experiments.
For example, suppose a certain data group is repeated 8 times in the production run of 50 trillion tokens. In a small 5-trillion-token experiment, that data group is scaled down to one-tenth and then repeated 8 times in the same way. The experiment does not simply cut the token count to one-tenth and feed the model new, different data one after another. This brings not only the model and data scale but also the condition of learning the same information many times closer to production.
Training efficiency when data is repeated has long been studied. The paper "Scaling Data-Constrained Language Models" by Niklas Muennighoff and colleagues varied compute budgets and data repetition counts and reported that the more times data is repeated, the smaller the improvement obtained from additional compute. That research differs from MAI-Base-2 in scale and conditions, so its repetition guidelines cannot be applied directly. Even so, it illustrates an important problem: processing the same number of tokens does not mean learning the same amount of new information.
According to Microsoft, if data-repetition conditions are not matched in small experiments, there is a risk of choosing settings that invite overfitting when scaled up. It therefore searched multiple combinations of learning rate and weight decay, validated the rules of thumb from small experiments in mid-scale experiments, and only then applied them to the production run. The efficiency gain was not achieved through any single new formula or algorithm. Reproducing production training conditions as closely as possible from the small-experiment stage, and thereby improving the accuracy of setting selection, was also an important factor.
MoE communication optimization and Muon improve training speed
Placing MoE in every layer was advantageous in terms of computation in Microsoft's past experiments too. However, the previous-generation MAI-Base-1 used alternating dense FFN and MoE layers. Even if the theoretical computation is smaller, processing on real GPUs slows down because of the burden of communication and memory management. Including these implementation costs, the alternating layout let the previous generation train in less time.
One major challenge with MoE is that the number of tokens assigned to each expert network is not uniform. If memory is reserved assuming the worst case, where the most tokens concentrate on one expert, large areas go unused in normal operation. On the other hand, in Microsoft's experiments, discarding tokens that exceed processing capacity degraded model performance. What was needed was a mechanism that processes all necessary tokens without wasting compute resources.
MAI-Base-2 fixes the processing capacity of each expert network and sends tokens that exceed that capacity to an additional processing round. It also shares spare capacity within the same group to cope with load concentrating on particular experts. In addition, it overlaps the transfer of data between GPUs, the actual computation, and the collection of results, reducing waste from waiting on communication. It also cuts the overhead of launching many small operations and of transferring data between GPUs. With these improvements, compute utilization in the MoE layers rose from 24% to 35%, making it possible to adopt a configuration with MoE in every layer.
For the optimization algorithm used to update weights, it adopted "Muon." Muon itself was not newly developed by Microsoft; the paper "Muon is Scalable for LLM Training" by Jingyuan Liu and colleagues also reports results from adding weight decay and update-scale adjustments and applying it to MoE model training. What Microsoft worked on was making this method run efficiently in an extremely large distributed training environment.
Muon performs matrix operations on the update history of weight matrices, so information stored split across multiple GPUs has to be gathered, computed on, and redistributed. This burden grows as models get larger, so Microsoft added optimizations such as handling matrices of the same shape together and running communication and computation in parallel. As a result, in the final training configuration, Muon's processing time was held to 2.1% of the average time of one training iteration. However, the efficiency gain is the result of combining multiple techniques, and Microsoft's 4x-to-12x multiples cannot be attributed to Muon alone.
To stabilize training, it also introduced a technique called "activation gradient clipping." When extremely large gradients arise with respect to the model's internal activations, this mechanism suppresses their influence locally. With ordinary gradient clipping, a small number of anomalous data points can weaken the update for the entire batch; controlling at a finer granularity is intended to limit that effect. Microsoft says this mechanism also helped it discover a defect in code data, where preprocessing had cut sentences off midway. An improvement to the training algorithm thus also helped identify a problem in data processing.
Actual pretraining of MAI-Base-2 used 16,000 NVIDIA GB200 GPUs from August to September 2026, with computation running for 38 days. MFU (Model FLOPs Utilization), which expresses model compute utilization over the whole run, was 21%, and the share of time during which training was progressing was 89%. MFU is calculated against the GB200's theoretical BF16 performance, and operations from recomputation and optimizer processing are not included in the numerator. The 89%, meanwhile, is the share of time training was actually progressing, not a statement that 89% of the GPUs' computing capacity was used. Because it also measures something different from the 35% MoE-layer compute utilization mentioned above, these figures cannot be compared as the same metric.
This effort shows the importance not just of reducing theoretical computation but of improving GPU communication, memory management, and optimizer processing to shorten actual training time. However, even if pretraining compute efficiency improves, inference speed, memory usage, and API pricing of the finished model will not necessarily improve by the same proportion. Efficiency at training time and efficiency when users actually run the model need to be evaluated separately.
For enterprise use, the focus is evaluation after post-training
Microsoft says that this pretraining did not use distillation from other companies' model outputs and that the tokens used for training were processed in-house. It emphasizes being able to trace data provenance, and this stance is tied to its plan to license model weights to trusted partners through Microsoft Foundry. Beyond simply developing its own foundation model, the company appears to be building an environment in which enterprises can tailor and operate models for their purposes.
However, the announcement does not mean MAI-Base-2's weights will be released for anyone to obtain freely. The model is currently in post-training, which shapes capabilities such as instruction following and reasoning, and neither an official availability date nor pricing has been disclosed. Nor can the statement that distillation was not used in pretraining be assumed to apply to all later training stages. The performance and efficiency figures published so far concern the model at the pretraining stage only.
To decide whether to adopt MAI-Base-2, enterprises will need to confirm how well the post-trained model performs in actual work. Independent evaluations comparing it with competing models under the same conditions will matter especially for code generation and modification, chart reading, and specialized reasoning. Response speed, pricing, and the license terms for using the model weights will also influence adoption. The results demonstrate progress in techniques for training large AI models efficiently, but how much performance improvement or cost reduction they bring to users will have to wait for the release of the finished model and verification in real-world operation.