Microsoft AI says it has completed pre-training MAI-Base-2, a mixture-of-experts (MoE) foundation model with 2.8 trillion total parameters and 80 billion active parameters per token. In the company’s October 15, 2026 announcement, Microsoft reports that the model beat every open-weight model it tested on held-out evaluations and was estimated to be 4.2–12.3 times more pre-training-FLOPs-efficient than DeepSeek-V4-Pro-Base.
What Microsoft announced
MAI-Base-2 is a pre-trained foundation model rather than a finished public chatbot or API. Its MoE design stores 2.8 trillion parameters, but routes each token through only about 80 billion active parameters. That sparsity lets Microsoft build a much larger model without executing the full parameter set for every token.
As an Amazon Associate I earn from qualifying purchases.
The announcement comes from Microsoft AI on October 15, 2026. The company says MAI-Base-2 was still undergoing post-training, so the reported pre-training results should not be treated as the final behavior of a released assistant.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →How the 2.8T total and 80B active figures work
Total parameters
The 2.8 trillion figure counts all learned weights across the model, including experts that are not selected for every token. It describes storage and overall model capacity, not the amount of computation used on each forward pass.
#1 Best Overall
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
Active parameters
Only 80 billion parameters are active per token under the announced routing scheme. In practical terms, different tokens can be sent to different expert subnetworks while sharing the model’s common components. Active-parameter counts are therefore more useful than total counts when estimating per-token compute, although memory, communication and serving requirements still depend on the full expert pool and system design.
What “4.2–12.3× more FLOPs-efficient” means
Microsoft’s efficiency claim is an estimate, not an independently reproduced benchmark. The company fitted a separate scaling law relating bits-per-byte to training floating-point operations for each of four held-out evaluations. It then estimated how much compute its own training recipe would need to reach the performance attributed to DeepSeek-V4-Pro-Base. Across those evaluations, the resulting estimate ranged from 4.2× to 12.3× fewer pre-training FLOPs for an equivalent result.
Rank #2
That methodology measures the efficiency of a training recipe and model scaling trend; it does not mean MAI-Base-2 will answer requests 4–12 times faster, cost 4–12 times less to serve, or outperform every model in every downstream application. Microsoft’s comparison is specifically against DeepSeek-V4-Pro-Base and is based on fitted pre-training curves.
What Microsoft measured
For held-out text, Microsoft used bits-per-byte (BPB), a tokenizer-agnostic metric. The company says MAI-Base-2 achieved lower BPB than the open-weight models included in its tests. Code was evaluated on Microsoft’s own codebase. STEM, mathematics and question-answering evaluations were commissioned from vendors, and Microsoft says the sets were decontaminated against the training data.
The comparison set included DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. Microsoft cautions that post-training can affect the results for the latter three models, so their placement in a pre-training comparison is not a clean measure of their final deployed capabilities.
Training scale, data and modality
| Item | MAI-Base-2 announcement |
|---|---|
| Training hardware | 16,000 GB200 GPUs |
| Training duration | 38 days during August and September 2026 |
| Average MFU | 21% |
| Goodput | 89% |
| Training mixture | 50 trillion tokens |
| Unique vision-language tokens | 4.5 trillion |
| Vision-language share | 18% of the mixture |
| Knowledge cutoff | March 2026 |
Microsoft says the pre-training data was organic and did not involve distillation from other models. The mixture’s vision-language component means the base model was trained beyond text alone, although the announcement does not establish the exact multimodal input and output features that will be exposed after post-training.
Rank #4
Engineering choices behind the result
Microsoft attributes the reported gains to three connected efforts: scaling ladders that include source scaling, co-design between the model and its infrastructure, and larger data pipelines that add vision information.
MoE efficiency
The company reports improving MoE-layer model-flop utilization (MFU) from 24% to 35%. That improvement made it practical to use MoE layers throughout the network rather than only in selected blocks. Microsoft also describes multi-round dropless MoE routing, intended to avoid dropping tokens when expert capacity is full.
Best Value
Optimization and stability
The training recipe used the Muon optimizer and activation-gradient clipping. These choices are presented as part of the system that allowed a very large sparse model to train at the stated scale; the announcement does not isolate how much each technique contributed.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How MAI-Base-2 compares with DeepSeek-V4-Pro-Base
| Comparison point | What Microsoft reports for MAI-Base-2 | What the announcement establishes about DeepSeek-V4-Pro-Base |
|---|---|---|
| Parameters | 2.8 trillion total; 80 billion active per token | Total and active counts are not stated in the announcement |
| Held-out text metric | Lower BPB than the open-weight models tested, according to Microsoft | Used as the principal open-model efficiency reference |
| Pre-training efficiency | Estimated 4.2–12.3× more FLOPs-efficient, based on four fitted scaling laws | Baseline for Microsoft’s fitted compute comparison |
| Evaluation status | Held-out sets; decontamination reported by Microsoft | Included in the same comparison framework |
Because the result comes from Microsoft’s own evaluations and scaling-law fit, readers should distinguish a company-reported advantage from an independently verified leaderboard result. Differences in data mixtures, code tasks, evaluation vendors and post-training can change the apparent gap.
Can you download or run MAI-Base-2?
Not at publication. Microsoft says the model was undergoing post-training and that it plans to license weights to trusted partners through Microsoft Foundry. The announcement gives no public weights, inference pricing or general release date, and it does not confirm availability for individual developers. Any Foundry access should therefore be treated as a future enterprise route rather than a currently open download.
Recommended Free Tools
Quick Recap
What to watch next
- Post-training evaluations that show how the base model performs after instruction tuning and alignment.
- Independent checks of the held-out BPB, code, STEM, mathematics and question-answering claims.
- Details of the licensing terms, supported Foundry regions and hardware requirements for trusted partners.
- Measured serving costs and latency, which cannot be inferred directly from the 4.2–12.3× pre-training estimate.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
