Microsoft discloses MAI-Base-2 pretraining: 2.8T total params, 38 days on 16k GB200s

Microsoft AI published a “Lab Notes” post on October 15 laying out the pretraining of its second in-house foundation model, MAI-Base-2: a mixture-of-experts base model with 80 billion active parameters out of 2.8 trillion total, trained for 38 days on 16,000 GB200 GPUs between August and September 2026. Closed labs rarely open the ledger this far — the post carries MFU, goodput, a training-recipe table, the scaling methodology and every held-out evaluation number. It also carries the limits that not shipping the weights buys them.

Key points

• MAI-Base-2 is an 80B-active / 2.8T-total MoE trained on 50T tokens for 38 days on 16,000 GB200s, at an average MFU of 21% and goodput of 89% (microsoft.ai, 2026-10-15).
• Microsoft claims lower bits-per-byte than every open-weight model it tested on four held-out sets, and 4.2–12.3x more pretraining FLOPs efficiency than DeepSeek-V4-Pro-Base and 6.3–12.7x more than MAI-Base-1.
• The weights are not released. The stated purpose is licensing to “trusted partners” through Foundry; the model is still in post-training, and preparations for MAI-Base-3 are already underway.

Background

This is not a from-scratch story. MAI-Base-1 shipped in June 2026 under MAI-Thinking-1, the reasoning model Microsoft AI put at roughly 52.8% on SWE-Bench Pro and 97% on AIME 2025, with the company insisting that batch of seven models was trained from scratch without distilling third-party models. Crypto Briefing records Base-1’s ledger: 30T tokens of pretraining data, 3.55T tokens of mid-training, 8,192 GB200s.

The hardware line runs in parallel. Our October 13 entry on Microsoft’s Maia 300 inference chip covered multi-gigawatt deployments starting in 2027 — and this post’s “model-infrastructure co-design” argument sits next to it: keep the recipe, the chips and the racks in step.

The comparison set says as much about the market as the numbers do. DeepSeek-V4-Pro-Base is the only released base pretraining checkpoint above one trillion parameters, which is why it became the efficiency reference — and it is also the model that Nex-N2.5, the open-weight model that topped BrowseComp, was post-trained from (1.6T total, 49B active). Several recent open-weight releases were built on the same ground.

The facts

Every architecture and training figure below comes from the official post or its PDF twin. All numbers are Microsoft’s own.

Item	MAI-Base-1	MAI-Base-2
Active / total params	35B / 1T	80B / 2.8T
Layers	78	102
Hidden dimension	6656	8704
MoE layout	Alternating dense FFN and MoE	MoE in every layer
Routing / experts	0 shared + 8/512 routed	1 shared + 7/512 routed
Attention heads (KV / Q)	8/80 global (every 6 layers) and 8/80 local	2/96 global (every 6 layers) and 16/96 local
Embeddings	Tied	Untied
Training horizon	30T tokens	50T tokens
Optimizer	AdamW	Muon (non-embedding 2D params) + AdamW
FP8 scheme	Tensor-wise, delayed scaling E4M3 / E5M2	MXFP8 E4M3
Vision	Text only	Native VLM

The run itself: 16,000 GB200s, 38 days, spanning August and September 2026, average MFU 21%, goodput 89%. Microsoft defines both terms in the post — MFU is normalized by 2.5 PFLOPS (GB200’s theoretical maximum in BF16) and excludes FLOPs for rematerialization and the optimizer; goodput is the fraction of time training made progress.

Evaluation uses bits-per-byte, which Microsoft argues is agnostic to each model’s tokenizer. The comparison covers frontier checkpoints available before October 2026 with at least 1T total parameters:

Model	Active/total	Code	QA	STEM	Math
MAI-Base-2	80B / 2.8T	0.2203	0.4162	0.3653	0.3561
Kimi K3	104B / 2.8T	0.2264	0.4269	0.3761	0.3589
Qwen3.8-2.4T-A95B	95B / 2.4T	0.2370	0.4351	0.3968	0.3752
DeepSeek-V4-Pro-Base	49B / 1.6T	0.2401	0.4427	0.3978	0.3803
MiMo-V2.6-Pro-RL	42B / 1.02T	0.2446	0.4440	0.4026	0.3862
MAI-Base-1	35B / 1T	0.2400	0.4513	0.3975	0.3939

Four caveats travel with that table. The held-out code set is Microsoft’s own codebase. The STEM, math and Q&A sets were commissioned from vendors to reduce contamination risk, and all evaluation sets were decontaminated against the training data. Kimi K3, Qwen3.8 and MiMo exist only as post-trained checkpoints, and Microsoft flags that post-training may affect those results.

The data recipe is the most concrete part of the disclosure. The corpus grew roughly 36% over MAI-Base-1 in unique training tokens, web data has a March 2026 knowledge cutoff, every token was processed in-house, and the crawler respects robots.txt plus related meta-tag and HTML controls. The mix is 45% code, 12% web, 9% PDFs, 9% STEM, 5% books, 4% math, 14% vision-only and 2% other. The VLM corpus holds 4.5T unique tokens; with repetition, vision data was 18% of the 50T-token mixture. The investing.com piece Yahoo Finance carried repeats these numbers but adds no independent sourcing, so they are treated here as relay, not verification.

Scaling method

The parts worth copying into your own recipe are the scaling ladder and what Microsoft calls source scaling. A ladder is a family of models on a fixed recipe; each rung is named after its layer count L, and L alone fixes the rest — hidden size is (256/3)·L, query heads are L rounded to the nearest multiple of 16. Models grow while tokens-per-active-parameter stays fixed. The problem is that small runs have to repeat data, and if the repeat distribution shifts, the small run picks architectures and hyperparameters that overfit at scale. Microsoft restricts unique samples on the lower rungs so every dataset repeats the same number of times regardless of horizon — a dataset seen 8x over the full 50T run is looped 8x over a tenth of the data in a 5T run. Under fixed TPP that holds the unique-data-to-parameters ratio constant, which the company credits for scaling reliably to 2.8T parameters.

Hyperparameters came out of a law rather than a grid. Learning rate and weight decay interact, so most of a rectangular sweep wastes samples; Microsoft swept the product of the two, fitted a 2D quadratic and took the analytical minimum. The scaling exponent of that product against steps sits close to −1, and LR’s against steps close to −0.5, matching the rule that doubling batch size should scale the learning rate by about √2.

MoE and optimizer work

MoE in every layer only became affordable once two implementation problems were fixed — dynamic routing causing memory fragmentation and reserving memory for extreme expert load imbalance, and per-expert overhead growing as experts got finer. Fixing them lifted MoE-layer MFU from 24% to 35%, which Microsoft credits for the architecture choice. Multi-round dropless MoE keeps memory static under any routing imbalance: experts are grouped, overflow tokens run in extra rounds, overloaded experts borrow unused capacity inside their group, and capacity is tuned so steady state needs a single round.

Muon’s overhead was cut by resharding momentum during the optimizer step, batching all-to-all by parameter shape and replicating rather than sharding router weights. After those changes Muon took 2.1% (165 ms) of the average training iteration, against a theoretical floor Microsoft estimates at 55 ms and attributes the gap to stragglers, kernel launch overhead, load imbalance and memory copies during reshardding.

What others say

Crypto Briefing skipped Base-2’s numbers entirely and worked the predecessor’s baseline and the open questions: whether Microsoft discloses a data and compute ledger to compare against 30T tokens and 8,192 GPUs, whether a reasoning model on Base-2 can clear the roughly 52.8% SWE-Bench Pro score, and whether the timing lines up with the 2027 Maia expansion. It also carries Microsoft’s own framing — no distillation, in-house data pipelines — which is the same narrative as the official post, not a second source.

investing.com, syndicated through Yahoo Finance, is the dense relay: all the numbers, no executive quotes, no third-party verification. It pushed 4.2–12.3x and 6.3–12.7x, the 36% corpus growth and the 18% vision share into the financial press.

Hacker News concentrated on the comparison set (27 points, 3 comments as of October 16). adroitboss first asked whether the model is open at all, then asked why a closed model is compared against open weights; leothetechguy answered that SOTA closed base models are not available to compare against. Bolwin flagged two disclosure gaps: the architecture looks very similar to Kimi K3 but with fewer active parameters (80B against 104B), and “I don’t think they mentioned what form of attention they use.” On the last point this entry checked the source: the post gives KV and Q head counts plus the global/local pattern but never names the attention mechanism, so the complaint holds.

Our take

For teams training their own base model, this is an unusually granular public ledger. Source scaling’s bookkeeping, sweeping hyperparameters by scaling law, dropless MoE’s memory ceiling and Muon’s resharding at 16k GPUs are four things you can test against your own recipe directly. MFU and goodput also come with their normalization stated, which makes them far easier to compare than the usual marketing numbers.

The edges are equally clear. Every efficiency and BPB claim is self-reported with no independent replication; the comparison set is open-weight only; three of the four held-out sets are vendor-commissioned and the code set is Microsoft’s own repository; and three of the five rivals are post-trained checkpoints. The model is still in post-training, the weights are not released, and there is no price and no date. For the audience this is aimed at — enterprises weighing a Foundry license — the post is where you start asking what a license would cost and what it would let you keep, not a model you can download today.

Read it as an unusually good pretraining report rather than a model launch.