Microsoft Says Its New Base Model Beats DeepSeek by 12x on Efficiency. The Baseline Is Six Months Old.

Microsoft AI published a dense technical write-up on October 15 introducing MAI-Base-2, a new foundation model it describes as state-of-the-art among open and proprietary base models. The headline number: Microsoft says the model needed 4.2 to 12.3 times fewer pretraining FLOPs than DeepSeek-V4-Pro-Base to reach equivalent performance, measured across four held-out evaluations. For a company that has spent the past year building an in-house model stack to reduce its dependence on OpenAI and Anthropic, a double-digit efficiency multiple over the most-cited open-weight base model is the kind of claim meant to get attention. It deserves scrutiny instead of a restatement, because the comparison Microsoft built it on is less symmetric than the headline suggests.

What Microsoft actually built

MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token, trained on a 50-trillion-token mixture (about 36% more unique tokens than its predecessor, MAI-Base-1) across 16,000 GB200 GPUs over 38 days in August and September. Architecturally, Microsoft moved to MoE layers throughout the network rather than alternating MoE with dense feed-forward blocks, which the team says lifted MoE model FLOPs utilization from 24% to 35%. It also switched its optimizer to Muon, which Microsoft reports added only about 2.1% to average iteration time against a theoretical peak of 55 milliseconds, and added selective activation-gradient clipping that it says grows more useful at scale. The model was also pretrained natively on vision data using an encoder-free architecture, a first for the MAI-Base line.

Those are real, falsifiable engineering choices, and Microsoft is unusually specific about them, down to the hyperparameter scaling-law formulas it used to set learning rate and dropout by model size. The efficiency claim is where the specificity gets slippery.

An efficiency claim measured against an old target

Microsoft's method was to fit independent scaling laws of bits-per-byte (a tokenizer-agnostic, lower-is-better measure) against training FLOPs, using DeepSeek-V4-Pro-Base as the reference point, then read off how many fewer FLOPs MAI-Base-2 needed to hit the same BPB. Microsoft's own post notes that DeepSeek-V4-Pro-Base was trained on "significantly fewer FLOPs" than MAI-Base-2 — which is true, but also somewhat beside the point if the question is which lab's current pretraining recipe is more efficient today. DeepSeek-V4-Pro-Base, according to independent model-tracker DataLearner, shipped on April 24, 2026: a 1.6-trillion-parameter, 49-billion-active MoE with a knowledge cutoff in mid-2025. That is roughly six months and, by DeepSeek's own public release cadence, at least one architecture generation behind the September 2026 training run Microsoft is benchmarking against it.

The other three named comparison models complicate the picture further. Microsoft's post states, in its own methods section, that Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL were only available to it as post-trained checkpoints — not as raw base models. Post-training routinely changes a model's behavior on held-out evaluations in ways pretraining-FLOPs comparisons aren't built to isolate. Microsoft discloses the caveat, then plots all four rivals on the same bits-per-byte chart anyway.

MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations.

That sentence is accurate to what Microsoft measured. It is not the same claim as beating DeepSeek's current pretraining recipe, and the two are easy to conflate if a reader stops at the 4-to-12x figure.

A claim that sits awkwardly next to Microsoft's own recent record

The timing raises a second question. Two months before MAI-Base-2, Microsoft shipped MAI Code 1.1 Flash, a coding model built on the same MAI-Base family, for GitHub Copilot. Independent testing by The Decoder found it lost to DeepSeek-V4-Flash-0731 by roughly 20 points on Terminal-Bench 2.1, while costing more than DeepSeek on every published pricing tier — input, cached input, and output tokens alike. Microsoft's own announcement for that model leaned on relative metrics (25% more token-efficient and a quarter the cost of its June predecessor) rather than head-to-head numbers against DeepSeek, which only showed up in the model card.

Pretraining FLOPs-efficiency and downstream task performance are genuinely different measurements, and a base model's advantage can be erased or reversed by post-training, serving infrastructure, or product-level tuning choices. But if MAI-Base-2 really represents the step-change in pretraining efficiency Microsoft describes, the gap between that claim and a shipped Microsoft product still losing to DeepSeek on price and benchmarks as recently as August is exactly the kind of discrepancy practitioners should track once MAI-Base-2's post-trained descendants actually ship.

What would settle it

Right now, nobody outside Microsoft can run MAI-Base-2 against a current model. Weights are going only to "trusted partners" through Microsoft Foundry, not to the public the way DeepSeek, Qwen, and Kimi publish their checkpoints — so there is no way to independently reproduce Microsoft's bits-per-byte numbers or test the model against anything DeepSeek has shipped since April. Microsoft says post-training is still underway and that more results are coming "soon." Until an independently-weighed, contemporaneous comparison exists, the 4-to-12x figure is best read as Microsoft's internal extrapolation against a six-month-old open checkpoint — a real signal about MAI-Base-2's architecture and training recipe, but not yet a verified capability lead over what DeepSeek, Qwen, or Moonshot are training right now.