MAI-Base-2: Inside Microsoft's 2.8 Trillion-Parameter Model
Microsoft finished pre-training MAI-Base-2, a 2.8 trillion-parameter mixture-of-experts model. Here is how it was trained, what the efficiency claims mean, and what businesses should do about it.
The Short Answer
On October 15, 2026, Microsoft AI announced that it had finished pre-training MAI-Base-2, a foundation model with 2.8 trillion total parameters and 80 billion active parameters per token. It is the largest base model Microsoft has ever built, and the company says it beat every open-weight model it tested on its own held-out evaluations. The…
On October 15, 2026, Microsoft AI announced that it had finished pre-training MAI-Base-2, a foundation model with 2.8 trillion total parameters and 80 billion active parameters per token. It is the largest base model Microsoft has ever built, and the company says it beat every open-weight model it tested on its own held-out evaluations. The model is not available to use yet. It is now in post-training, the stage where a raw base model is taught to follow instructions, reason, and behave safely.
This article explains what MAI-Base-2 is, how Microsoft trained it, what the headline efficiency claims really mean, and what businesses should take away from it.
What Is MAI-Base-2?
MAI-Base-2 is a mixture-of-experts (MoE) foundation model. A foundation model is a large AI model trained on vast amounts of data that can later be adapted for many different tasks. A base model is the raw result of pre-training: it has learned to predict text and understand patterns, but it has not yet been taught to chat, follow instructions, or reason step by step.
MAI-Base-2 is the successor to MAI-Base-1, which Microsoft announced in June 2026. MAI-Base-1 became the foundation for MAI-Thinking-1, Microsoft's first frontier reasoning model. MAI-Base-2 is larger and sparser than its predecessor: far more total parameters, but still only a small fraction active for any given token.
The mixture-of-experts design is the key to that trade-off. Instead of running every parameter on every input, an MoE model routes each token to a subset of specialized "expert" sub-networks. Microsoft says MAI-Base-2 has expert layers throughout the model, in every layer, made possible by improving how efficiently those expert layers run.
For businesses exploring AI and machine learning services, the important point is that base models like MAI-Base-2 are the raw material of the AI products that arrive later.
MAI-Base-2 by the Numbers
| Detail | MAI-Base-2 |
|---|---|
| Architecture | Mixture-of-experts foundation model, with expert layers in every layer |
| Total parameters | 2.8 trillion |
| Active parameters per token | 80 billion (about 2.9 percent of the total) |
| Training hardware | 16,000 Nvidia GB200 GPUs |
| Training time | 38 days, across August and September 2026 |
| Training data | About 50 trillion tokens, with no distillation from other models |
| Unique token growth vs MAI-Base-1 | About 36 percent more unique training tokens |
| Vision data share | 18 percent of the training mix, learned natively from the start |
| Knowledge cutoff | March 2026 |
| Training efficiency | 21 percent average model FLOPs utilization, 89 percent goodput |
| Status | In post-training, not released |
| Planned access | Weights to be licensed to trusted partners through Foundry |
How Microsoft Trained MAI-Base-2
Microsoft trained MAI-Base-2 on 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026, processing a training mixture of about 50 trillion tokens. Two utilization figures from the run tell a story about how hard it is to train at this scale. Average model FLOPs utilization reached 21 percent, a measure of how much of the GPUs' peak computing speed was actually put to work. Goodput reached 89 percent, meaning the cluster spent 89 percent of its time making useful training progress, with the rest lost to hardware failures, network interruptions, and checkpoint overhead.
Microsoft also improved the efficiency of the MoE expert layers themselves, raising their compute utilization from 24 percent to 35 percent compared with earlier approaches. That fix is what made it practical to put expert layers in every layer of the model.
The training recipe also included improvements to the Muon optimizer, a method for updating a model's weights during training, and a technique Microsoft calls source scaling, designed to make small-scale experiments reliably predict large-scale behavior. By artificially restricting unique training samples at smaller scales, each experiment repeats its dataset the same number of times regardless of training length, which prevents small experiments from favoring choices that would overfit at full scale.
The headline result of all this work is efficiency. Microsoft estimates that its training recipe is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which it describes as the strongest open base model, and 6.3 to 12.7 times more efficient than its own MAI-Base-1 from June. Those figures are estimates derived by fitting scaling laws to results on four held-out evaluations, not direct head-to-head training runs.
Native Vision From Day One
One of the biggest differences between MAI-Base-2 and its predecessor is vision. MAI-Base-1 was a text model. MAI-Base-2 was pre-trained natively with vision, meaning it learned images and text together from the start, with no separate image encoder bolted on afterward. Vision-language data made up 18 percent of the training mixture.
Native multimodal training matters because so much real business work is visual: charts in board decks, diagrams in manuals, screenshots in bug reports, scans in medical files, and photos on factory floors. A model that learned vision alongside text from the beginning can, in principle, reason across both in a more unified way than a model that had vision grafted on later.
The data behind the model is another point Microsoft emphasized. The company says its pre-training is built on "organic" data, that it processed every training token in-house, and that it involved no distillation from other companies' models. Distillation is the practice of training a model on the outputs of another model. Microsoft also says its proprietary web crawler respects robots.txt and related controls. Clean data provenance matters for a business reason: it is what allows Microsoft to license the model's weights to partners, as explained below.
MAI-Base-2 and the Efficiency Claim: What 4x to 12x Really Means
Microsoft's claim that MAI-Base-2 is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base is the most quoted number in the announcement, and also the easiest to misread. It is worth being precise about what it does and does not say.
First, it measures estimated pre-training efficiency, not inference speed, operating cost, or intelligence. It says Microsoft believes it can reach a given evaluation score using far less training compute than the comparison model needed. It does not say the model answers questions faster, costs less to run, or is smarter.
Second, the comparison rests on Microsoft's own evaluations. Microsoft says MAI-Base-2 outperformed every open-weight model it tested on held-out evaluations, using bits-per-byte on held-out text, including code from Microsoft AI's own codebase and vendor-commissioned STEM, math, and question-answering evaluations designed to reduce training-set contamination. Those results have not been independently replicated. Microsoft also acknowledged that for some comparison models, only post-trained checkpoints were available, and that post-training may affect the evaluations.
Third, even taken at face value, the claim is about a recipe, not just a model. If Microsoft's scaling science genuinely lets it train frontier models with far less compute, that advantage compounds across every future model it builds, including MAI-Base-3, which Microsoft says is already underway.
MAI-Base-2 vs Maia 300: Do Not Mix Them Up
Microsoft made two big AI announcements in the same week in October 2026, and their similar names cause real confusion. They are completely different things.
MAI-Base-2 is a software model: an AI foundation model defined by its weights, its training data, and its architecture. It was trained on Nvidia GB200 GPUs, and it is currently in post-training.
Maia 300 is hardware: Microsoft's third-generation custom AI chip, announced on October 12, 2026. It is designed for inference, which means running AI models after they are trained, not training them. Microsoft says Maia 300 will become the primary inference system for its products and services, with large-scale deployments starting in 2027.
The relationship between the two is strategic, not technical. Microsoft is building its own models and its own chips at the same time, so it can tune them together and reduce its dependence on outside suppliers. You can read more about the chip in our earlier article on what Maia 300 means for cloud strategy.
What MAI-Base-2 Means for Businesses
MAI-Base-2 is not a product any business can buy or use today. Its importance is as a signal of where Microsoft's AI platform is heading.
1. Microsoft is becoming a model maker, not just a model reseller
For years, Microsoft's flagship AI offerings were closely tied to OpenAI's models. MAI-Base-2 shows Microsoft building an in-house model pipeline of its own. That gives Microsoft more control over cost, capability, and roadmap, and it gives customers a second frontier model family inside the Microsoft ecosystem.
2. Licensed weights could open new enterprise options
Microsoft says it plans to license MAI-Base-2 weights to trusted partners through Foundry. That is different from the usual cloud model offer, where you call an API and never see the model itself. Licensed weights could let large enterprises and partners run, fine-tune, and deeply integrate a frontier-class model inside their own systems. No date, pricing, or partner list has been announced, so treat this as a direction to plan around, not a product to buy yet.
3. Efficiency gains could lower the cost of AI features
If Microsoft's efficiency claims hold up, the cost of training and serving frontier models falls, which eventually shows up as lower prices or better performance for the AI features businesses pay for. For a sense of how fast the model landscape is moving, compare this announcement with other recent launches covered on this blog, including Google's Gemini 4 Argon and Meta's Muse Spark open model.
4. Plan for multimodal AI now
MAI-Base-2's native vision is part of a wider shift: frontier models are becoming multimodal by default. Businesses should assume that the next generation of AI tools will read charts, screenshots, forms, and photos as fluently as they read text. Teams that need help turning this shift into a practical plan can start with IT consulting to map which workflows benefit first from multimodal models.
5. Do not rebuild your roadmap on a base model
The most common mistake after an announcement like this is to pause real projects while waiting for the new model. MAI-Base-2 is in post-training with no release date. The right move is to keep building on models available today, with architectures that make it easy to swap in better models later.
What We Still Do Not Know
Honest coverage of MAI-Base-2 has to include the gaps. Microsoft has not announced a release date, pricing, or partner list for the licensed weights. It has not published independent benchmark results, and its comparisons have not been replicated by third parties. We do not know how the post-trained version will perform on coding, agentic work, or long-context reasoning, which are the tasks businesses actually care about.
There is also a wider question about Microsoft's OpenAI relationship. Microsoft holds a large stake in OpenAI and sells its models on Azure, while now saying its own base model is ahead on its internal tests. How the two companies' partnership evolves as Microsoft's in-house models improve is one of the most important stories in AI to watch.
FAQ About MAI-Base-2
What is MAI-Base-2?
MAI-Base-2 is Microsoft's newest foundation model, announced on October 15, 2026. It is a mixture-of-experts base model with 2.8 trillion total parameters and 80 billion active parameters per token. It has completed pre-training and is now in post-training.
Can I use MAI-Base-2 today?
No. MAI-Base-2 is not publicly available. It is undergoing post-training, and Microsoft has not announced a release date. The company plans to license the model's weights to trusted partners through Foundry, but no date, pricing, or partner list has been shared yet.
How big is MAI-Base-2 compared with other models?
At 2.8 trillion total parameters, it is one of the largest publicly confirmed parameter counts from any lab, and the largest in Microsoft's own model family. Because it is a mixture-of-experts model, only about 80 billion parameters are active for each token, which keeps the cost of running it far below what a dense model of that size would require.
Is MAI-Base-2 the same as Maia 300?
No. MAI-Base-2 is an AI software model. Maia 300 is a custom AI chip designed for inference, announced on October 12, 2026, with large-scale deployments planned from 2027. MAI-Base-2 was trained on Nvidia GB200 GPUs, not on Maia chips.
What does Microsoft claim about MAI-Base-2's efficiency?
Microsoft estimates its training recipe is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, and 6.3 to 12.7 times more efficient than its own MAI-Base-1. These are Microsoft's own scaling-law estimates based on held-out evaluations, not independently verified results, and they measure training efficiency rather than speed, cost, or intelligence.
What should businesses do about MAI-Base-2 now?
Keep building on models available today, design systems so models can be swapped in later, and plan for multimodal AI as the default. Watch for post-training results, Foundry licensing details, and the first products built on MAI-Base-2.
MAI-Base-2 is a research milestone, not a product launch, but it is a milestone that points clearly in one direction: Microsoft intends to own more of the AI stack, from data and training recipes to models and chips. Businesses that understand that direction early will be better placed to take advantage of it when MAI-based products arrive. If you want help preparing your AI roadmap for the next generation of models, contact our team and we will map out the practical next steps for your business.
Related Articles
Keep reading — more guides you may find useful.
Related Services
Services connected to this guide.
