Markdown edition

# Microsoft’s MAI-Base-2 report measures pretraining, not a released model

> Microsoft reports MAI-Base-2's held-out text results and fitted training-efficiency estimates. The base model is still in post-training, with no public checkpoint to test.

By BIG CHANGE Editorial

Published: 2026-10-16T16:13:14.938Z
Updated: 2026-10-16T16:13:14.938Z
Canonical: https://bigchange.ai/blog/microsoft-mai-base-2-pretraining-efficiency-report

![Conceptual charcoal illustration of a storage rack connected by overhead cables to compute racks, with a separate blank evaluation workstation in a generic room.](https://bigchange.ai/api/media/file/mai-base-2-pretraining-concept-v1.png)
AI-generated conceptual illustration by BIG CHANGE; not actual Microsoft hardware or a measured result.

Microsoft AI says it finished pretraining MAI-Base-2, a mixture-of-experts model with 80 billion active parameters and 2.8 trillion total parameters. Its [October 15 technical report](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/) describes a 38-day run on 16,000 GB200 GPUs in August and September 2026. For engineers assessing the result, the useful question is narrower than a model ranking: what does the report measure before post-training and public access?

The answer has two parts. Microsoft reports lower bits per byte (BPB) on four held-out evaluation categories than the checkpoints it tested. It separately estimates that its training recipe needs 4.2–12.3 times fewer training FLOPs than DeepSeek-V4-Pro-Base to reach matched BPB, using curves fitted to its own scaling ladder. Those are different findings with different limits.

## The big change

- **What changed:** Microsoft has reported a completed large-scale base-model run, with held-out text evaluations and a detailed account of its training recipe. MAI-Base-2 remains in post-training.
- **Why it matters:** Engineers can examine how Microsoft relates model quality to training compute, and which data, architecture and infrastructure choices it says contributed. The evidence concerns pretraining loss and Microsoft's internal measurements; there is no public MAI-Base-2 inference service or released checkpoint to test against a workload.
- **What to watch:** Post-training results and access terms will determine whether this foundation becomes useful in products or to external teams. Microsoft's stated plan to license weights through Foundry is a future offering, not present availability.

## What the held-out comparison says

Microsoft evaluates pre-trained models on held-out text with BPB, which expresses prediction loss per byte and avoids a tokenizer-dependent score. The report names code, STEM, math and Q&A evaluations. It says the code set comes from MAI's own codebase, while Microsoft commissioned the other sets from vendors; Microsoft says it decontaminated all four against its training data. These are safeguards described by the evaluator, not an independently reproduced audit of the data.

On those sets, Microsoft reports that MAI-Base-2 has lower BPB than every model it tested. The checkpoint distinction matters. DeepSeek-V4-Pro-Base is the released *base* model Microsoft uses as its recipe reference. Microsoft says only *post-trained* checkpoints were available for Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, and acknowledges that post-training can affect these evaluations. A base-versus-post-trained BPB chart is therefore not a controlled comparison of pretraining recipes for all four rivals. The report gives no OpenAI checkpoint test, so it supplies no basis for a claim that MAI-Base-2 beats an OpenAI model.

BPB also answers a different question from assistant usefulness. It measures loss on the selected held-out material, not instruction following, reasoning after post-training, latency, safety or operating cost for an application. Microsoft says MAI-Base-2 is still undergoing post-training; its report does not provide public weights, an inference endpoint, pricing or customer outcomes.

## Where the compute-efficiency range comes from

The 4.2–12.3× claim comes from fitted scaling curves. Microsoft fit an independent BPB-versus-training-FLOPs curve for each of the four evaluations using its scaling ladder and the known final MAI-Base-2 result. It then estimated how much compute its recipe would need to match DeepSeek-V4-Pro-Base’s BPB and compared that estimate with the reference model’s reported training compute. Microsoft says the matching points lie within its fitted model range, rather than being extrapolated beyond it.

The range is a *matched-loss estimate* for Microsoft's recipe on those evaluations. It is not a measured reduction in MAI-Base-2's actual training bill, a speedup for inference or proof that another lab would obtain the same result. Microsoft also reports a 6.3–12.7× estimated FLOPs-efficiency improvement against its own MAI-Base-1. In both cases, the reader needs the chosen reference, evaluation set and fitting method to interpret the number.

Microsoft gives another pair of numbers for the production run: average model FLOPs utilization (MFU) of 21% and goodput of 89%, where goodput is the fraction of time training progressed. The report normalizes MFU to a theoretical 2.5 PFLOPS per GB200 in BF16 and excludes rematerialization and optimizer FLOPs. MFU concerns how much of that theoretical compute capacity the training computation used; goodput concerns interruption and progress over elapsed time. Neither is the matched-BPB FLOPs-efficiency ratio. Microsoft explicitly distinguishes gains measured in training FLOPs from gains measured in wall-clock time when it evaluates design choices.

## The recipe Microsoft says produced the result

The report offers plausible mechanisms, along with the limits of its evidence. Its scaling ladder restricts unique samples on smaller runs so datasets repeat the same number of times as in the full run. Microsoft says that prevented small experiments from favoring choices that would overfit at larger scale. It also describes empirical learning-rate, weight-decay and dropout scaling rules, checked at an intermediate model size before the full run.

In the sparse model, Microsoft says changes to expert routing, fixed memory capacity and grouped communication raised MFU for its MoE layers from 24% to 35%, allowing MoE in every layer. That is a component-level internal measurement, not the 21% average MFU for the entire 38-day run. The report describes optimizer work that left Muon at 165 milliseconds, or 2.1% of an average training iteration, plus activation-gradient clipping to stabilize training. Its data section says the unique-token corpus grew roughly 36% from MAI-Base-1. It reports 4.5 trillion unique vision-language tokens; with repetition, vision-language data made up 18% of the 50-trillion-token training mixture.

These details make the paper useful as an engineering account of how Microsoft approached scale. They do not isolate the causal contribution of each change across independently replicated runs. Microsoft attributes the result to data quality and modeling choices; external teams cannot establish those effects from the report alone. The immediate evidence is a completed base model, Microsoft’s selected held-out losses and its fitted recipe-efficiency analysis. Product performance remains a separate measurement for a later checkpoint.

## Sources & further reading

- [Microsoft AI, “Frontier Pre-training at Microsoft AI”](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/), October 15, 2026. First-party technical report for architecture, training run, held-out BPB comparisons, fitted compute-efficiency ranges, utilization, data and stated release stage. All measurements and implementation effects here are Microsoft-reported; BIG CHANGE did not replicate the training or evaluate MAI-Base-2.

## Sources

- [Microsoft AI: Frontier Pre-training at Microsoft AI](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/) — First-party technical report; source of all quantitative claims and implementation descriptions. Performance and mechanisms are Microsoft-reported, not independently replicated.