Microsoft AI has completed pre-training on MAI-Base-2, a new mixture-of-experts foundation model with 80 billion active parameters and 2.8 trillion total parameters, the division announced October 15 in a technical "Lab Notes" post. The company says the model outperforms every open-weight model it tested on held-out evaluations and is 4 to 12 times more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base.
The announcement is the clearest signal yet that Microsoft's in-house model program — built under CEO Mustafa Suleyman's Microsoft AI organization — intends to compete at the frontier rather than rely solely on its OpenAI partnership. MAI-Base-2 is now undergoing post-training, with more results promised soon. For ongoing coverage of new model releases, follow our AI model news.
The Scale of the Run
Microsoft says it trained MAI-Base-2 on 16,000 NVIDIA GB200 GPUs over 38 days in August and September 2026. The run averaged 21% model FLOPs utilization (normalized against the GB200's theoretical maximum of 2.5 PFLOPS in BF16, excluding rematerialization and optimizer FLOPs) and achieved 89% goodput — the fraction of time training progressed without interruption. Those numbers describe a disciplined large-scale run; interrupted or poorly utilized GPU fleets are the norm at this scale, and every point of goodput translates into weeks of saved training time.
The model is larger, sparser, and trained on more data than its predecessor, MAI-Base-1, which served as the foundation for MAI-Thinking-1, Microsoft's reasoning model released in June. Microsoft says it deliberately reinvested the efficiency savings from MAI-Base-2's training recipe into training a larger model.
How Microsoft Says It Compares
The company evaluated MAI-Base-2 against frontier checkpoints available before October 2026 using bits-per-byte on held-out text — a tokenizer-agnostic metric — along with held-out code evaluations drawn from Microsoft's own codebase. STEM, math, and Q&A evaluations were commissioned from external vendors to reduce the risk of training-set contamination, and all evaluation sets were decontaminated against the training data.
The comparison set included DeepSeek-V4-Pro-Base, described by Microsoft as the most capable base model with open weights and the only released base pre-training checkpoint above one trillion parameters, plus Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL. Microsoft notes a caveat: for those three, only post-trained checkpoints are available, and post-training may affect the pre-training evaluations. Across the held-out evaluations, Microsoft reports MAI-Base-2 achieved lower bits-per-byte than the other models tested.
The Efficiency Claim, Decoded
The 4-12x efficiency figure is not a direct benchmark win but a compute-matching analysis. For each held-out evaluation, Microsoft fitted a scaling law of bits-per-byte against training FLOPs and estimated how much compute its own recipe would need to match DeepSeek-V4-Pro-Base's score. The range reflects differences across the four evaluations. The claim, in plain terms: Microsoft says its recipe extracts several times more capability per unit of training compute than the strongest openly available base model — and it spent that surplus on scale.
Owning the Full Stack
Beyond raw performance, the post lays out the strategic rationale. Microsoft says owning its model pipeline end-to-end gives it strong data provenance over training tokens, and that pre-training relies on organic data with no distillation from other models. That provenance, the company argues, will let it license model weights to enterprises through Foundry as a trusted source of frontier intelligence, while co-tuning models and infrastructure lowers cost first in Microsoft's own products.
The technical section credits three initiatives for the run's success: scaling science, including "scaling ladders" that account for repeated training data via a technique called source scaling, which keeps the ratio of unique data to parameters fixed across rungs; model-infrastructure co-design, including improved MoE and Muon optimizer implementations alongside activation gradient clipping; and expanded data pipelines with the integration of vision data. Microsoft says source scaling was key to scaling reliably to the 2.8-trillion-parameter size without overfitting surprises.
What It Means for the Market
If the claims hold up under independent scrutiny, MAI-Base-2 positions Microsoft as a fourth major Western frontier-model builder alongside OpenAI, Anthropic, and Google — with a differentiated pitch built on data provenance and enterprise licensing rather than consumer chatbot mindshare. The market that matters first is Azure's enterprise customer base, where Foundry customers could gain access to a frontier-class model whose training lineage Microsoft fully controls.
Independent verification remains the open question. Microsoft's evaluations are self-reported, the comparison models' post-training caveat is material, and bits-per-byte on held-out text does not directly translate to the chat, coding, and agentic benchmarks enterprises actually shop with. Post-trained derivatives of MAI-Base-2 — the successors to MAI-Thinking-1 — will be the real test. But the underlying message of the announcement is unambiguous: Microsoft's bet on training frontier models in-house, on its own silicon fleet, is no longer aspirational. It has a finished base model, and the post-training pipeline is already running.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
