Microsoft AI says it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active, trained on 16,000 GB200 GPUs over 38 days. In its Lab Notes post Frontier Pre-training at Microsoft AI (October 15, 2026), the company claims the model "outperforms every open-weight model we tested on our held-out evaluations" and that the run was roughly 4 to 12 times more FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base.
Those are striking claims, and they rest on Microsoft's own evaluations. This post lays out the recipe, the numbers, what is checkable today, and what must wait for post-training results.
TL;DR: what is claimed and what is verified
| Item | Claim from Microsoft AI | Status |
|---|---|---|
| Model | MAI-Base-2, MoE, 2.8T total, 80B active, natively vision (encoder-free) | Described; no weights or public API |
| Training run | 16k GB200s, 38 days, Aug to Sep 2026, 21% MFU, 89% goodput | Self-reported |
| Data | 50T-token mix, 18% vision (4.5T unique vision tokens), cutoff March 2026 | Self-reported |
| Quality | Beats every open-weight model tested on held-out evals | Microsoft's own evals only |
| Efficiency | 4.2 to 12.3x vs DeepSeek-V4-Pro-Base; 6.3 to 12.7x vs MAI-Base-1 | Estimated from fitted scaling laws |
| Data provenance | "Organic," no distillation from other models | Not independently verified |
| Availability | Weights licensable to partners via Foundry; post-training under way | No release date |
What is MAI-Base-2?
It is a base model, meaning the raw output of pre-training before instruction tuning or reinforcement learning. It succeeds MAI-Base-1, the previous base model that Microsoft uses as its efficiency baseline. For Microsoft's earlier reasoning model, see our coverage of MAI-Thinking-1 in Foundry public preview. Microsoft says preparations for MAI-Base-3 are already underway.
MAI-Base-2 is larger and sparser than its predecessor: 80B parameters active per token out of 2.8T total. It has vision built in rather than bolted on through a separate image encoder. A mixture-of-experts design activates only a few expert sub-networks for each token, which is how a 2.8T model can run with the per-token compute of an 80B one.
How was it trained?
A grid of GPU chips with one row lit, illustrating the 16,000-GPU MAI-Base-2 pre-training run
The post is unusually technical for a corporate blog. The main ingredients:
- Scaling ladders. Microsoft trains models of increasing size with a fixed recipe and constant tokens per active parameter. A technique it calls "source scaling" limits the unique data in small runs so conclusions transfer to large ones.
- Hyperparameter scaling laws. Learning rate and weight decay were swept jointly, and Microsoft derived formulas for these and for dropout. It says the laws beat earlier heuristics at an intermediate scale.
- Mixture of experts in every layer. MoE MFU rose from 24% to 35% through changes including a multi-round "dropless" routing scheme with fixed memory use and lower per-expert overhead.
- Muon optimizer instead of AdamW. (Muon reference implementation.) Chosen for stability and efficiency. It costs 2.1% of iteration time (165 ms), against an estimated 55 ms theoretical floor, so there is still headroom.
- Activation gradient clipping. It clips gradients at selected activations rather than using the global norm. Microsoft says it also used the method to find a preprocessing bug that truncated a license header in code data.
Its guiding principle, quoted from the post: "Simple methods scale best."
What went into the data?
The corpus grew about 36% in unique tokens over MAI-Base-1. Microsoft cites quality classification, deduplication, decontamination and embedding-based diversity sampling. Web data comes from a proprietary crawler that respects robots.txt, and sources that violate Microsoft's Responsible AI policies or appear on the USTR Notorious Markets list are excluded. The vision corpus has 4.5T unique tokens and makes up 18% of the 50T-token mix.
The key provenance statement is that training data is "organic" and not distilled from other models. Microsoft ties that to its ability to license the weights to partners through Foundry. That is a commercial argument as much as a technical one: distilled data can carry the terms of the model it came from. The claim is plausible for a company with its own crawler, but the post offers no audit, so treat it as a statement, not a finding. For context on how Foundry packages Microsoft's models, see Microsoft Foundry naming explained.
What do the utilization numbers say about the run?
Two figures in the post are easy to skim past. Average model FLOPs utilization (MFU) was 21%, meaning roughly a fifth of the GPUs' theoretical arithmetic throughput turned into useful training math. That sounds low, but sparse mixture-of-experts models are communication-heavy because tokens must be routed between experts spread over many GPUs, and published MFU for large MoE runs is commonly well below that of dense models. Microsoft's own progress is the more telling number: it says MoE-layer MFU rose from 24% to 35% through its routing changes.
The second figure is goodput of 89%: the share of wall-clock time spent making training progress rather than recovering from failures, restarting or waiting. At 16,000 GPUs, hardware faults are routine, so a high goodput over 38 days points to solid fault handling. Together, 21% MFU and 89% goodput suggest the efficiency story is as much about not wasting compute as about clever architecture, and that further gains remain, as the Muon overhead gap shows.
A rough sanity check on scale: the post gives a 50T-token mix and 80B active parameters. The usual rule of thumb of six FLOPs per active parameter per token implies on the order of 2.4e25 training FLOPs. That is our back-of-envelope arithmetic, not a number Microsoft published, and it ignores attention and routing overhead, so treat it only as a sense of magnitude.
How does Microsoft measure the 4-12x efficiency gain?
This is the claim most likely to be quoted, so it deserves care.
The metric is bits-per-byte on held-out data, which is tokenizer-agnostic and lets Microsoft compare models with different vocabularies. For each benchmark, Microsoft fits a scaling law to estimate how much compute a reference model would need to match MAI-Base-2's score. Dividing gives an efficiency multiple. The result is a range (4.2 to 12.3x against DeepSeek-V4-Pro-Base) because different benchmarks give different multiples.
Competitors named in the comparison are DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. Code is evaluated on Microsoft's internal codebase. STEM, math and Q&A evaluations were commissioned from vendors and decontaminated against training data.
Reasons to read this carefully:
- Held-out means held out by Microsoft. Outsiders cannot rerun the suite.
- Extrapolated scaling laws. Efficiency multiples depend on fitted curves, and the range itself is wide.
- Mixed checkpoints. The post acknowledges that for several rivals it compared post-trained checkpoints, which may affect results.
- No downstream results. Nothing yet on instruction-following, coding agents or reasoning after post-training, which is where users feel quality.
A base model that wins on bits-per-byte is a strong foundation, not a finished product.
Why does this matter?
For a company that has leaned on OpenAI models, an in-house frontier-scale pre-train is a strategic marker. Microsoft AI has been shipping its own models for months: Microsoft-Decision-1, MAI-Code-1.1 Flash running locally, and its own silicon in the Maia 300 inference chip plan. A strong in-house base model is the piece that lets Microsoft control cost and licensing across all three.
It also feeds the open-weight race. Rival releases include Mistral Large 4 preview and Reflection AI's open-weight push. Microsoft's claim is that it can now match or beat the best open base models at a fraction of the compute, which, if true, would reset expectations about how much a frontier pre-train must cost. The Mozilla analysis of the open-weight frontier gap is a useful companion for judging how quickly such leads erode.
What this means for builders
- Nothing to download. There is no public checkpoint. If you build on Azure, watch Foundry for post-trained MAI models rather than the base.
- Licensing angle. If "organic data" holds up, enterprises that worry about inherited license terms from distilled models get a cleaner story. Ask Microsoft for contractual language, not just a blog claim.
- Cost expectations. If a 4-12x efficiency improvement is real, expect it to show up as price pressure on Microsoft's own endpoints over time, but there is no pricing in this post.
- Wait for post-training. Judge the model on tool use, coding and reasoning evaluations from independent testers, not on base-model loss curves.
What to watch next
Microsoft says more results are coming "soon." The signals that would raise confidence: post-training benchmarks that outside groups can reproduce, a technical report with the scaling-law fits, and any third-party check of the organic-data claim. Signals that would lower it: no public access for months, or independent testers failing to see the gap on tasks outside Microsoft's suite.
The post names no individual authors and describes a "lean, fast-moving lab" that is hiring. A Hacker News thread on the post drew only a handful of points when we checked, so independent reaction is thin so far.
Figures are as published by Microsoft AI on October 15, 2026. They are the company's claims and may be revised.