Microsoft's MAI-Base-2: 2.8T Params, 4–12× More Efficient Than DeepSeek-V4 (Microsoft's Claim)

Microsoft AI published a frontier pre-training blog today describing MAI-Base-2 — a 2.8T-parameter Mixture-of-Experts with 80B active parameters per token, trained on 50T tokens across 16,000 GB200s in 38 days. Microsoft claims 4–12× better pre-training FLOPs-efficiency than DeepSeek-V4-Pro-Base. Post-training is still underway. The model ships only to trusted partners via Foundry, not as open weights and not as a general API.

The blog is mostly about the training infrastructure, not the model capability. That is the interesting part. Microsoft is publishing enough of the recipe to claim a frontier training capability without exposing the model itself, and the recipe is the lever they want to demonstrate.

The Numbers

• 2.8T total parameters
• 80B active per token
• 50T training tokens
• 16,000 GB200 GPUs
• 38 days wall-clock
• Claimed 4–12× better pre-training FLOPs-efficiency than DeepSeek-V4-Pro-Base

For comparison: Llama 3 405B is 405B dense, trained on 15.6T tokens. DeepSeek-V3 is 671B/37B, trained on 14.8T tokens. MAI-Base-2 is bigger in total parameters, larger in active budget per token, and trained on a token count that is 3.4× the Llama 3 405B run. The cluster is also larger than any previously published training run from a non-Meta frontier lab.

The 38-day wall-clock is the operational headline. 50T tokens on 16,000 GB200s in 38 days implies an MFU in the 40–50% range, which is at the high end of what has been publicly reported for MoE training. Microsoft's blog discusses the optimizations that drove the MFU: Muon optimizer scaling, activation gradient clipping, custom communication kernels for the MoE all-to-all, and a cluster topology tuned for the workload.

Why This Matters

Microsoft is now in the frontier training conversation. With MAI-Base-2, Microsoft joins Meta, Google DeepMind, and the leading Chinese labs in demonstrating frontier-scale training capability on a model it intends to productize. The 2.8T scale, 50T tokens, and 16,000-GPU cluster are the operational credentials. The 4–12× efficiency claim is the technical credential. Both are necessary; neither is sufficient on its own.

The Muon optimizer is having a moment. MAI-Base-2 is the largest public training run to use a Muon-family optimizer at scale. The blog credits the optimizer with a substantial share of the FLOPs-efficiency claim. Muon was developed in part by Microsoft researchers and has been adopted in several open releases this year. If the MAI-Base-2 numbers hold up under independent benchmarking, expect Muon-family optimizers to become a default in frontier training rather than a research curiosity.

Foundry is the route, not the API. Microsoft is positioning MAI-Base-2 as a Foundry-exclusive model, not a general API. That is a meaningful change from the Azure-OpenAI pattern of "frontier model on Azure, billed by token." Foundry is the model-customization service for enterprises that want to fine-tune and deploy on dedicated capacity. The MAI-Base-2 release is the first frontier model positioned exclusively for that path.

The DeepSeek comparison is strategic. Microsoft chose DeepSeek-V4-Pro-Base as the comparison point, not Llama 3 405B or Qwen 3 Max. That is the competitor Microsoft sees itself relative to. The framing tells you who Microsoft thinks the open-weights frontier belongs to right now.

The Trade-Off

The 4–12× efficiency claim is Microsoft's. Independent verification has not yet been published. The training recipe is plausible at the scale described, and Microsoft's research organization has a track record of credible efficiency claims, but until a third party runs the same recipe and reproduces the number, treat the 4–12× as a vendor claim, not an audited benchmark.

The "no open weights, no general API" posture is also a real choice. The model exists, the training infrastructure exists, and the recipe is described in enough detail to be evaluable — but the model itself is locked to Foundry customers. Microsoft is keeping the lever (recipe) and the artifact (weights) separate, which gives it more strategic control than a pure open release would.

The Bottom Line

Microsoft is now in the frontier training conversation with a 2.8T-parameter MoE trained on 50T tokens in 38 days. The 4–12× efficiency claim is credible but vendor-attributed; wait for independent verification. The Muon optimizer at frontier scale is the operational story. Foundry-only access is the strategic choice. Builders who care about frontier training infrastructure: read the blog for the recipe, not for a model you can use today.