On Thursday, 15 October 2026, Microsoft AI, the in-house lab Mustafa Suleyman runs separately from Microsoft’s OpenAI partnership, published a lab note titled Frontier Pre-training at Microsoft AI. Its main announcement: the lab has finished pre-training MAI-Base-2, a mixture-of-experts (MoE) base model with 2.8 trillion total parameters and 80 billion active per token. Microsoft trained it on 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026.
MAI-Base-2 is not a product yet. It is a base model, meaning raw pre-trained weights with no instruction tuning and no reasoning RL, and Microsoft says it is “currently undergoing post-training.” You can’t call it from an API today. Even so, it is the most detailed account Microsoft has given of a frontier training run, and it tells you a lot about where the company wants to stand once it no longer relies only on OpenAI.
What was actually announced
According to the primary post and the coverage that cites it (GIGAZINE, OODAloop):
- Size: 2.8T total parameters, 80B active. Its predecessor MAI-Base-1, which underpins the MAI-Thinking-1 reasoning model Microsoft released in June 2026, was about 1T total and 35B active according to GIGAZINE.
- Data: a 50-trillion-token training mixture. The vision-language portion holds 4.5T unique tokens and makes up 18% of the mixture once repetition is counted. The unique-token corpus is about 36% larger than MAI-Base-1’s. Knowledge cutoff is March 2026, and the post says the crawler respects
robots.txt. - Modality: “pre-trained natively with vision using an encoder-free VLM architecture.” Images go straight into the transformer instead of through a separate vision encoder such as a CLIP-style tower.
- Provenance: “our path to frontier performance is built on organic data, not distillation from other models.”
- Access: Microsoft says it will “license our model weights to trusted partners through Foundry.” That makes the weights licensed to approved partners, not open-weight. No terms, prices or dates have been announced.
The engineering details worth a developer’s attention
Most of the note is about systems work, and that part is the most concrete:
- MoE in every layer. MAI-Base-1 used MoE layers only in every other layer, which the lab blames on “implementation inefficiencies.” MAI-Base-2 uses them throughout, with finer-grained experts and a “multi-round dropless MoE” that keeps memory use fixed. Model FLOPs utilisation (MFU) in the MoE layers rose from 24% to 35%.
- The Muon optimizer at scale. The lab uses Muon, the orthogonalising optimizer that has replaced AdamW in several recent frontier runs, with a sharded Newton–Schulz implementation. They report it at 2.1% of iteration time (165 ms), against a theoretical best of about 55 ms.
- Whole-run efficiency. Average MFU was 21%, measured against the GB200’s 2.5 PFLOPS BF16 peak, and goodput was 89%, meaning 89% of wall-clock time went to useful training rather than restarts and failures. At 16,000 GPUs that is a reasonable result. It is roughly in line with the ~90% reportedly achieved for MAI-Base-1 on 8,192 GPUs, according to summaries of the MAI-Thinking-1 technical report.
- A scaling ladder. Microsoft trained smaller models on a “ladder” where each rung is named after its layer count L, with hidden size (256/3)·L and the query-head count set to L rounded to a multiple of 16. That fixed geometry lets them fit clean scaling laws.
A back-of-envelope check shows the numbers agree with each other. 16,000 GPUs × 2.5 PFLOPS × 21% × 38 days comes to about 2.8 × 10²⁵ FLOPs. The standard 6·N·D estimate with 80B active parameters and 50T tokens gives about 2.4 × 10²⁵. Microsoft didn’t publish a total compute figure; this is my arithmetic.
How “4–12x more efficient” is calculated
The headline claim is that MAI-Base-2 is 4–12x (4.2–12.3x in the body) more pre-training-FLOPs-efficient than DeepSeek-V4-Pro-Base, which Microsoft calls the strongest open base model. Against MAI-Base-1 the gain is 6.3–12.7x.
The method matters here. Microsoft doesn’t compare final MMLU-style scores. It uses four held-out evaluation sets (code from Microsoft’s own codebase, plus STEM, math and Q&A sets commissioned from vendors and decontaminated against training data) and scores them in bits per byte (BPB), a measure of how well the model predicts unseen text. For each set it fits a curve of BPB against compute on its ladder models and asks how much compute its recipe would need to reach DeepSeek-V4-Pro-Base’s BPB. It then compares that figure with DeepSeek’s reported training compute. The range comes from the four sets disagreeing.
The approach is legitimate, but you should know what it measures:
- It is extrapolated. The 4–12x describes Microsoft’s recipe, inferred from fitted scaling curves. It doesn’t come from a head-to-head of two finished models on the same task suite.
- BPB isn’t capability. Lower perplexity on held-out text correlates with downstream quality, but it doesn’t guarantee better agentic coding or tool use after post-training.
- The evals aren’t public. One set is Microsoft’s internal code, and the others were commissioned from vendors. Nobody outside can reproduce the numbers.
- The rest of the comparison is post-trained. The note also compares against Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, but only post-trained checkpoints of those exist, so they aren’t strictly base-to-base. The figures appear only as charts, and none of the coverage I found reproduced the exact numbers.
GIGAZINE summarised the claim as Microsoft cutting pre-training compute “to one-twelfth” to match DeepSeek. That is the top of the range, not the typical result.
Why it matters
There are three reasons this matters beyond Microsoft’s own roadmap.
Microsoft is building a full stack of its own. A lab that can run 16,000 GB200s at 89% goodput, write its own MoE kernels and optimizer sharding, and crawl its own data has a credible fallback to OpenAI. The note says MAI-Base-3 preparation is “already underway.” For the enterprises that buy model access through Azure, Foundry may soon offer a first-party frontier model next to OpenAI’s.
Efficiency is now the main competitive claim. Reflection AI’s Beam a fortnight ago also led with efficiency rather than leaderboards, and now Microsoft does the same. When training budgets reach tens of millions of GPU-hours, a 4x recipe improvement is worth more than a few benchmark points. Expect more labs to publish scaling-ladder BPB comparisons, and to argue about them.
“No distillation” has become a selling point. After a year of accusations that labs train on each other’s outputs, Microsoft states that its data is organic and processed in-house. That matters for enterprise licensing, where data provenance and IP exposure are compliance questions.
Open questions
Several basics are still missing: expert count and routing, context length, attention design, any downstream benchmark of a post-trained MAI model built on this base, licensing terms, and a release date. The efficiency figure uses Microsoft’s own evals and Microsoft’s reading of DeepSeek’s compute. Until a post-trained model ships and independent evaluators can test it, MAI-Base-2 is a well-documented training run, not a proven product.