Microsoft’s MAI-Base-2: A 2.8-Trillion-Parameter Bet on AI Independence From OpenAI

Last updated: 16 October 2026. Figures in this piece are sourced to primary and industry publications, linked inline; where a number is self-reported by Microsoft rather than independently verified, that is flagged explicitly.

The 60-second version

• Microsoft AI says it has finished pretraining MAI-Base-2, a new in-house foundation model with 2.8 trillion total parameters and 80 billion active parameters, built as a mixture-of-experts (MoE) architecture.
• Microsoft claims the model “outperformed every open-weight model evaluated” and needed up to 12 times less compute to match China’s DeepSeek-V4-Pro-Base, the open-weight model it is benchmarking against.
• This is a pretraining milestone, not a product launch. MAI-Base-2 has not shipped in Copilot, Azure AI Foundry, or any public API, and Microsoft has not said whether — or when — it will.
• The release lands inside a wider strategy, run by Mustafa Suleyman’s Microsoft AI division, to reduce the company’s dependence on OpenAI’s models for its own products.
• Every efficiency and benchmark figure in Microsoft’s announcement is self-reported. No independent lab has reproduced the comparison yet.

What Microsoft actually announced

On 16 October, Microsoft AI said it had completed pretraining on MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters, of which 80 billion are active on any given token. That puts it in the same scale bracket as the largest MoE systems from Chinese open-weight labs, while keeping the active-parameter count — the part that actually determines inference cost — closer to a mid-sized dense model.

Microsoft’s own framing, repeated across the outlets that covered the announcement, is twofold: first, that MAI-Base-2 “outperformed every open-weight model evaluated” on the benchmarks Microsoft ran; second, that it did so while requiring roughly four to twelve times less compute than DeepSeek-V4-Pro-Base needs to reach comparable results — Microsoft’s own post puts the precise range at 4.2x to 12.3x. Microsoft’s post also states the training run directly: roughly 50 trillion tokens, on about 16,000 Nvidia GB200 GPUs, over 38 days in August and September 2026. Those figures come from Microsoft’s own announcement, not a secondary account, and are corroborated by multiple outlets that covered it.

What Microsoft has not said is just as telling. There is no release date, no pricing, and no confirmation of open weights. Microsoft says it plans to eventually license the model’s weights to trusted partners through Azure AI Foundry, but has given no timeline for that and no word on whether MAI-Base-2 will power Copilot or a public API. “Completed pretraining” is a research-pipeline term: it means the expensive, compute-heavy phase is done, not that a usable, safety-tested, cost-optimised product exists yet. Microsoft’s own MAI-1-preview took months to go from a similar pretraining milestone to a shipping product.

Key numbers

Why this matters: Microsoft’s slow walk away from OpenAI

MAI-Base-2 is not Microsoft’s first in-house model, but it is by far its largest. The company has spent 2026 building out the Microsoft AI division under Mustafa Suleyman specifically to reduce reliance on OpenAI, whose models still sit under most of Copilot. That strategic context matters more than any single benchmark number: Microsoft remains OpenAI’s largest infrastructure partner and a major shareholder, even as it builds a parallel, fully-owned model stack it can route around that relationship whenever it wants to. The two positions are not contradictory so much as a hedge — and MAI-Base-2 is the biggest chip Microsoft has placed on that hedge so far.

It also arrives at a moment when the economics of frontier training are under real scrutiny. OpenAI’s own compute bill has become a story in its own right — its latest funding round valued the company at exactly the same figure as its compute commitments — and every major lab is now selling “efficiency” as aggressively as it sells raw capability. A training run that Microsoft says needed four to twelve times less compute than a rival’s flagship to match it is, in that climate, as much a cost argument aimed at Microsoft’s own board and Azure customers as it is a capability claim aimed at OpenAI or Google.

The DeepSeek comparison, and what “4–12x more efficient” really means

Efficiency claims of this kind are notoriously slippery, for three reasons. First, the range itself — four to twelve times — is wide enough to cover very different stories depending on which benchmark and which configuration produced which end of it; Microsoft has not published the full benchmark-by-benchmark breakdown alongside the headline figure. Second, “compute to match” depends entirely on what counts as a match: matching DeepSeek on a reasoning benchmark is a different exercise from matching it on code generation or long-context retrieval, and an aggregate efficiency number can hide large variance across tasks. Third, and most importantly, this is Microsoft grading its own model against a rival’s using benchmarks Microsoft selected and ran in-house. DeepSeek has not published a response, and no third-party lab has reproduced the comparison.

None of that makes the claim false. It does make it unverified, and worth treating the way any reasonable reader should treat a vendor’s own performance marketing: as a starting hypothesis, not a settled fact. The pattern is familiar from Reflection AI’s own 501-billion-parameter Beam model, launched this month explicitly to rival China’s open-weight releases — another case where the company’s own benchmarks did the talking before independent evaluation caught up. If MAI-Base-2 is ever released for third-party testing, expect the real efficiency number to land somewhere narrower than “4–12x” once outside labs run their own comparisons.

What most coverage is getting wrong — or leaving out

Most of the write-ups that circulated within hours of the announcement repeated the headline figures without flagging that this is a pretraining-complete milestone, not a shipping model — a distinction that matters enormously for anyone trying to work out what changes today versus what might change in six months. A completed pretraining run tells you a company can train at a given scale; it tells you almost nothing about inference cost, latency, safety testing, or whether the resulting model will ever reach a product a developer can actually call.

The coverage has also mostly skipped the compute-sourcing angle. Training a 2.8-trillion-parameter model at the scale reported requires GPU capacity that very few organisations on Earth can assemble, and Microsoft’s position as both a hyperscaler and a major AI lab gives it an advantage here that pure-play labs do not have. That advantage is the same one at the centre of Anthropic’s recent Nscale compute deal, where Google and Microsoft’s own exit from the arrangement said as much about the shifting compute market as the deal itself. Microsoft training its own frontier model on its own infrastructure, rather than buying capacity from someone else’s cluster, is a quietly significant data point in that same story.

Practical takeaways for people who build, publish, or run software

For teams making real decisions this week, rather than reacting to the headline, a few things follow directly from what is and isn’t confirmed:

• Don’t plan around MAI-Base-2 yet. There is no API, no pricing, and no confirmed release window. Anyone evaluating model options for a product shipping in the next two quarters should treat this as a “watch” item, not an input to the decision.
• If cost-per-token is your bottleneck, the real signal to track is the independent benchmark, not Microsoft’s. When (or if) third-party evaluators such as LMSYS, Artificial Analysis, or academic groups publish their own comparisons, that number — not the 4–12x range — is the one worth planning around.
• Multi-model strategies keep paying off. Teams already routing between providers, as discussed in our look at how open-source and proprietary AI really compare in practice, are better placed to absorb a new entrant like MAI-Base-2 than teams locked to a single vendor’s stack.
• Watch Azure pricing, not just Copilot features. If MAI-Base-2 ever ships, the more consequential move for most businesses will be where it appears in Azure AI Foundry’s price list relative to GPT and Gemini-class models, not whether Copilot’s marketing mentions it.

What we still don’t know

• The exact terms and timeline for releasing MAI-Base-2 — Microsoft says it plans to license weights to trusted partners via Azure AI Foundry, but hasn’t set a date, pricing, or said whether it will reach Copilot or a public API.
• The full benchmark breakdown behind the “4.2–12.3x” efficiency claim, including which specific tasks produced the low end and the high end of that range.
• How MAI-Base-2 performs against proprietary frontier models from OpenAI, Google, and Anthropic, none of which featured in Microsoft’s “every open-weight model evaluated” comparison.
• Whether this changes Microsoft’s commercial terms with OpenAI in any concrete way, or remains a parallel research track with no near-term effect on Copilot.

FAQ

Is MAI-Base-2 available to use right now?
No. Microsoft has only announced that pretraining is complete. There is no public API, no confirmed product integration, and no stated release date.

Is MAI-Base-2 actually better than DeepSeek’s model?
Microsoft says it outperformed every open-weight model it evaluated and needed far less compute than DeepSeek-V4-Pro-Base to match. Those are Microsoft’s own, self-reported benchmark results; they have not been independently verified.

Does this mean Microsoft is dropping OpenAI?
No. Microsoft remains deeply invested in OpenAI commercially and infrastructurally. MAI-Base-2 is best read as Microsoft building leverage and optionality, not as a divorce.

Will MAI-Base-2 be open-weight?
Unconfirmed. Microsoft has not said whether the model, if released, would ship with open or closed weights.

Sources

• Microsoft AI — “Frontier pre-training at Microsoft AI” (primary announcement)
• AI Weekly — AI News Today, 16 October 2026
• GIGAZINE — Microsoft announces completion of pre-training for MAI-Base-2
• TokenPost — Microsoft Completes Pretraining of 2.8 Trillion-Parameter MAI-Base-2
• FourWeekMBA — Microsoft Builds 2.8T-Parameter Model to Rival Open Weights
• BigGo Finance — Microsoft Unveils 2.8 Trillion-Parameter AI Model, Claiming 4–12x Efficiency Gains Over DeepSeek
• Outlook Business — Microsoft Unveils First In-House MAI Models: Bid for Independence From OpenAI

Please follow and like us: