MAI-Base-2 Finished Pre-Training: Microsoft's 4–12x Efficiency Claim, Minus the Model

The number that travelled on October 15 was "3T". MAI-Base-2, Microsoft AI's new in-house foundation model, is the model behind it — and the company's own technical post puts it at 2.8 trillion total parameters with 80 billion active. The rounding is minor; the more useful thing is what the same post does and does not contain. MAI-Base-2 has finished pre-training, is currently being post-trained, and has no checkpoint, no price, no licence and no date attached to it. The headline claim is a 4–12x pre-training FLOPs-efficiency advantage over DeepSeek-V4-Pro-Base, the only released base pretraining checkpoint above a trillion parameters. Everything about that claim is a vendor estimate, and the post is unusually explicit about how it was produced. This is what the public record supports as of October 16, 2026 — and the two numbers in it that Microsoft did not put in the headline.

What Microsoft actually published, and what it did not

The primary source is a Lab Notes post on Microsoft AI's own site, dated October 15, 2026, titled "Frontier Pre-training at Microsoft AI." Mustafa Suleyman announced it the same afternoon with the line that MAI now has "state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI," and closed with "post-training results coming soon." That is the whole basis of the story, and it is a company describing its own work — the technical claims below are unreproduced by anyone outside MAI, and there is no independent evaluation of this model in the public record.

What does not exist yet is the part most readers want:

• Weights — none. No repository, no checkpoint, no quantisation, no download of any kind.

• A product or an API — none. There is no endpoint, no pricing page, no model card.

• A licence — not stated. The post says owning the pipeline "will enable us to license our model weights to trusted partners through Foundry," which is a stated intention, not an announcement.

• A date — none for post-training results, none for Foundry availability, none for MAI-Base-3.

• Benchmarks anyone can rerun — none. The evaluation surface is described, and the raw scores are not published in a form a third party can reproduce.

So the accurate description of MAI-Base-2 today is one word: unfinished. Pre-training — the expensive, months-long phase that turns raw text and images into a base model — is done. Post-training is where a base model becomes something a person can talk to, and that phase is running now. The distinction matters because it is exactly where the interesting failure modes live, and because a base checkpoint and a chat model are not comparable objects.

What "4–12x more FLOPs-efficient" actually measures

This is the claim that will be repeated everywhere, so it is worth being precise about its shape, because it is not a head-to-head benchmark run.

The definition, from the post: Microsoft compares pre-trained models on held-out text using bits-per-byte, a tokenizer-agnostic metric — lower is better. Held-out code is drawn from MAI's own codebase; the STEM, Math and Q&A sets were commissioned from outside vendors "to reduce the risk of training-set contamination," and all evaluation sets were decontaminated against the training data. That is a more careful evaluation description than most launch posts carry, and it is still the company grading its own work on its own held-out sets.

The efficiency figure is then derived, not measured directly. For each of four held-out evaluations, MAI fitted a scaling law of bits-per-byte against training FLOPs across its internal model ladder, included MAI-Base-2's own final performance in the fits so that the matching point falls inside the fitted range rather than beyond it, and estimated how much compute a model on its ladder would need to match DeepSeek-V4-Pro-Base. Compared against that model's actual training compute, the answer is 4.2–12.3x more FLOPs-efficient; against MAI-Base-1, which shipped in June, the improvement is 6.3–12.7x. The "4–12x" of the headline is that first range, rounded.

Three things about this are worth holding on to:

• It is an estimate from a fitted curve, not two models trained under controlled conditions. The post is candid about this, including the detail that the reference model "is still trained on significantly fewer FLOPs than MAI-Base-2."

• The comparison set is narrow by rule. Only models with at least one trillion total parameters, and only frontier checkpoints available before October 2026, were eligible. DeepSeek-V4-Pro-Base is the sole released base pretraining checkpoint that qualified.

• Three of the names it is compared against only exist as post-trained checkpoints. Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL are reported with an explicit vendor caveat that post-training may have affected the evaluations. Comparing a fresh base checkpoint against post-trained models on a next-token-prediction metric is not a like-for-like contest, and Microsoft says so rather than letting the chart imply otherwise.

None of that makes the claim wrong. It makes it a claim — the kind a reader should treat as a direction until someone reproduces it.

The compute number Microsoft left out

The post gives everything needed to compute MAI-Base-2's training run, and then does not compute it. The run was 16,000 GB200 GPUs over 38 days in August and September 2026, at 21% average model FLOPs utilization, with the utilization normalized against GB200's theoretical 2.5 PFLOPS in BF16.

Multiply those out — 16,000 accelerators at 2.5 × 10¹⁵ FLOPs per second, times 38 days, times 0.21 — and the run implies roughly 2.8 × 10²⁵ FLOPs of training compute. That figure is our arithmetic from the post's own inputs, not a number Microsoft published. It is a lower bound, because the post's own definition of MFU excludes the FLOPs spent on rematerialization and on the optimizer. And it sits well above the 10²⁵ FLOPs threshold the EU AI Act uses as its presumption of systemic risk for a general-purpose AI model: roughly two to three times that line, from a single pre-training run.

To be clear about what that does and does not mean: no obligation attaches to a model nobody can access. The relevance is that the compute total is the one figure in the post that is a matter of arithmetic from published inputs rather than a fitted estimate, and when the model that inherits this base eventually ships — through Microsoft's own products or to Foundry partners — this is the number regulators will start from.

What is actually inside the model

Strip out the efficiency argument and the post is a fairly detailed account of three engineering problems, which is the part with lasting value regardless of who wins a benchmark.

• Scale — an MoE foundation model with 80 billion active and 2.8 trillion total parameters, larger and sparser than MAI-Base-1.

• Mixture of experts in every layer — MAI-Base-1 alternated dense feed-forward blocks with MoE blocks because its MoE implementation was not fast enough to justify more. Fixing dynamic-routing memory fragmentation and per-expert overhead, plus grouping computation and communication per expert group, raised MoE layer MFU from 24% to 35%, which is what made MoE in every layer the time-optimal choice.

• Dropless MoE with fixed memory — instead of dropping overflow tokens or over-allocating for the worst case, MAI groups experts and runs additional dispatch-compute-collect rounds for overflow, borrowing unused capacity from other experts in the same group. Memory usage is static under arbitrary routing imbalance; the post's argument against token dropping is that it measurably hurts quality.

• Muon at scale — the optimizer is resharded, cast to BF16, exchanged all-to-all, orthogonalized locally and returned. After optimization it accounted for 2.1% of the average training iteration, or 165 ms, against a theoretical peak of 55 ms that the post says is bottlenecked by communication. Muon was adopted over AdamW for more stable dynamics.

• Activation gradient clipping — clipping at selected activations rather than on the global batch norm, which kept gradient norms in a much tighter band than MAI-Base-1's, and the benefit grew with scale. It also turned out to be a data-quality diagnostic: a malformed training example with a licence header cut off mid-sentence, traced back through a code-data pull request, surfaced a preprocessing bug that has since been fixed.

• Data — a corpus about 36% larger than MAI-Base-1's in unique training tokens, feeding a 50-trillion-token mixture, with a March 2026 knowledge cutoff. The company says the pipeline is organic data with no distillation from other models, processed in-house, and that its proprietary crawler honours robots.txt and excludes sources on the USTR Notorious Markets list.

• Vision from the start — unlike MAI-Base-1, this model is pre-trained natively with vision through an encoder-free architecture. The vision-language corpus holds 4.5 trillion unique tokens and accounted for 18% of the 50-trillion-token mixture, spanning image captioning, chart understanding, visual grounding, spatial reasoning, medical understanding, computer use and sheet-music reading.

There are also the hyperparameter scaling laws, published in closed form: the optimal learning rate times weight decay is 17.04 divided by the number of training steps, the optimal learning rate follows a log-linear function of tokens per active parameter and steps, and dropout is a sigmoid of the logarithm of active parameters. Publishing those is the single most reusable thing in the post — a scaling ladder whose rungs are named by layer count alone, where hidden size is fixed at 256/3 times the layer count and query heads equal the layer count rounded to the nearest multiple of 16, is a recipe other labs can test.

Why a base checkpoint is not a product, and what happens next

MAI-Thinking-1, the reasoning model Microsoft introduced earlier this year, was post-trained on top of MAI-Base-1. The post says MAI-Base-2 is now "the foundation for our next generation of models," which means the model people will actually be able to use is not MAI-Base-2 — it is whatever Microsoft post-trains on top of it, and that name has not been announced.

The distribution channel is also already decided, and it is not open weights. The stated plan is to license weights "to trusted partners through Foundry," which describes an enterprise procurement route rather than a download. Nothing in the post promises public weights, and the emphasis on data provenance and the absence of distillation reads as groundwork for exactly that kind of licensed offer.

Then there is MAI-Base-3. The post's only sentence on it is that "preparations for MAI-Base-3 are already underway." No parameters, no compute, no data, no dates — a roadmap signal rather than information. What it does establish is cadence: MAI-Base-1 released in June, its successor's pre-training reported complete in mid-October, and the next cycle already in preparation. A base-model generation roughly every four months, at this scale, is the more consequential claim in the post than any single benchmark row.

What you can call today

Nothing in the MAI-Base-2 story is callable, and it is worth saying plainly that none of Microsoft's MAI models is on OrcaRouter's catalogue today — we do not host them, and a base checkpoint is not a served model in any case. What the post does hand you is a comparison set, and most of that set is reachable right now. The reference model in Microsoft's efficiency claim, the post-trained DeepSeek V4 Pro, is on our catalogue, as are Kimi K3 and Qwen3.8-Max, two of the open-weight models MAI-Base-2 was measured against.

That gives the claim a cheap test. One API across 200+ models means comparing what you can actually call does not require a second contract or a second SDK, and because OrcaRouter passes provider list prices through with 0% markup rather than marking them up, a price move by any of those vendors is live on our side the same day. If you want to run a workload across the post-trained versions of these models before deciding whether to wait for whatever Microsoft ships on top of MAI-Base-2, that is the comparison a base-model benchmark cannot give you — and the routing DSL lets you move traffic between them by configuration rather than code when the answer changes.

What would change this piece

Four things, in order of how much they would matter.

The first is post-training results. Everything in the October 15 post describes a pre-trained artifact; whether the model behind it is good at following instructions, calling tools or holding a long conversation is a separate question that has not been asked in public.

The second is any form of access — a Foundry listing with a price, a partner announcement, or weights with a licence attached. At that point this stops being a research report and becomes a procurement decision.

The third is independent measurement. Bits-per-byte on the vendor's own codebase, with vendor-commissioned STEM and math sets, is not something an outside lab can check today. A third party reproducing the scaling-law comparison against DeepSeek-V4-Pro-Base would settle the efficiency claim; until then it is a credible, well-documented, self-reported estimate.

The fourth is MAI-Base-3 detail. "Preparations are underway" is a sentence, not a spec, and the generation after next is where the four-month cadence either holds or does not.

The narrow accurate summary: Microsoft has finished pre-training its largest in-house model, published an unusually technical account of how, claimed an efficiency advantage over the best open base checkpoint, disclosed a training run of roughly 2.8 × 10²⁵ FLOPs by arithmetic it did not do itself in the post, and shipped nothing anyone can use.

Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.