Microsoft says MAI-Base-2 pre-training is complete

On 15 October 2026 Microsoft AI published Frontier Pre-training at Microsoft AI. The page carries that date. The PDF linked from the page names, in its metadata, the author as Microsoft AI. The post says pre-training of MAI-Base-2 is complete.

The post is two days old. The line that goes past it is a Business Insider report, timestamped 2026-10-15T23:31:44Z, of a remark Mustafa Suleyman, CEO of Microsoft AI, is said to have made to staff. The OpenAI comparison is in that report. It is not in the Microsoft post.

The run Microsoft describes

Microsoft calls MAI-Base-2 a mixture-of-experts foundation model with 80 billion active parameters and 2.8 trillion total parameters. It says the model was trained on 16,000 GB200 GPUs, written as “16k” in the post, over 38 days in August and September 2026. It says average model FLOPs utilization over the run was 21 percent and goodput, the share of time training progressed, was 89 percent. It says the utilization figure is normalized by 2.5 petaflops, which it calls the GB200’s theoretical maximum in BF16, and that the figure leaves out FLOPs for rematerialization and the optimizer. Every number in this paragraph is Microsoft’s description of its own run.

Microsoft says the expanded web pipeline has a March 2026 knowledge cutoff, and that vision data was 18 percent of a 50-trillion-token mixture. Those figures are Microsoft’s.

The efficiency claim Microsoft publishes

The opening says MAI-Base-2 is 4 to 12 times more efficient in pre-training FLOPs than DeepSeek-V4-Pro-Base, which Microsoft calls the strongest open base model. Later, Microsoft explains the comparison. For each of four held-out evaluations it fits a scaling law of bits per byte against training FLOPs, estimates how much compute its own recipe would need to match DeepSeek-V4-Pro-Base, and compares that estimate with DeepSeek’s training compute. It says MAI-Base-2’s final result is included in the fits. The result it states is that its recipe is 4.2 to 12.3 times more FLOPs-efficient. The same paragraph says progress since MAI-Base-1, which it places in June, is a 6.3 to 12.7 times FLOPs-efficiency gain. Microsoft attributes the gain to its data and its modeling choices.

Microsoft also says some models on its chart were available only as post-trained checkpoints: Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL. It says post-training may affect those results. The 4.2 to 12.3 times figure is the comparison with DeepSeek-V4-Pro-Base, which Microsoft presents as a base checkpoint. None of these scores is an independent benchmark. They are Microsoft’s measurements on Microsoft’s evaluations.

Still in post-training

The post says MAI-Base-2 is currently undergoing post-training, and that preparations for MAI-Base-3 are already underway. It does not announce an API, a public weight file, or a release date. It says owning the pipeline, and training on data it describes as organic rather than distilled from other models, “will enable” Microsoft to license model weights to trusted partners through Foundry. That sentence describes a license Microsoft says the setup will make possible. The post publishes no weights.

What Business Insider reports from a town hall

Business Insider marks the story as subscriber-only. The passages below are Ashley Stewart’s report of an internal meeting. AI Tamer did not hear the meeting, and Microsoft’s post does not repeat the comparison.

Stewart writes that Suleyman told employees on Thursday: “We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol.” The next sentence she quotes is: “And in fact, we used one-quarter of the compute that OpenAI used to train their model.” She writes that he said Microsoft had privately tested the model against OpenAI models, including one nicknamed “Spud” that powers GPT-5.6, and that Microsoft’s model did better on tests of how well AI understands and predicts patterns. She writes that he said the comparisons use technical information Microsoft has about OpenAI’s models, and she quotes him: “Privately and confidentially, we’ve also compared it to the OpenAI models.” She writes that Microsoft declined to comment and that OpenAI did not respond to a request for comment on Thursday. Thursday was 15 October 2026, the date on both the Microsoft post and the Insider timestamp.

Stewart also quotes Suleyman saying the new model used “12 times less computation than was used for DeepSeek, for example, on compute efficiency.” She writes that he did not name the DeepSeek model. The post’s published figure against DeepSeek-V4-Pro-Base is the range 4.2 to 12.3 times, from the scaling-law method above. “12 times less,” with no model named, is a different sentence.

The public record on 17 October is the post and the PDF: Microsoft’s account of a finished pre-training run for an 80-billion-active, 2.8-trillion-total mixture-of-experts model, still in post-training. The claim that this base model is stronger than the one behind GPT-5.6 Sol, at a quarter of the compute, is Business Insider’s report of what Suleyman told staff.