On October 15, 2026, Microsoft AI published an article in its official "Lab Notes" series, formally announcing the completion of MAI-Base-2 pre-training. The model is a mixture-of-experts (MoE) foundation model with a total parameter count of 2.8 trillion, activating 80 billion parameters per token, trained on 50 trillion tokens of corpus, 18% of which is visual data (4.5 trillion independent visual tokens in total), with a knowledge cutoff date of March 2026.
According to the official microsoft.ai blog, the training infrastructure was extremely large: 16,000 NVIDIA GB200 GPUs running in parallel for 38 days, spanning August to September 2026, with a measured model FLOPs utilization (MFU) of 21% and goodput (effective training time ratio) reaching 89%. Microsoft also said that preparations for MAI-Base-3 have already begun in parallel.
Behind the Efficiency Numbers: Scaling-Law Estimates, Not Direct Comparison
Microsoft claims that MAI-Base-2's pre-training FLOPs efficiency is 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base, and 6.3 to 12.7 times higher than its own MAI-Base-1. This is the most eye-catching number in this release, but it is also the one that most warrants caution.
According to explainx.ai's technical interpretation of the blog, the above efficiency gap was not directly measured from two actual training runs, but was estimated from scaling laws fitted by Microsoft itself. Microsoft's approach was: first train models at different scales, fit a curve of performance as a function of compute, and then extrapolate to reach efficiency comparison conclusions. This method has academic legitimacy, but it is fundamentally different from "the ratio of compute required to achieve the same performance in actual tests"—the former is a prediction, the latter is a fact.
All evaluation results likewise come from Microsoft's internal "held-out evaluations", and no third-party institution has independently reproduced them. The model weights are not currently public and are licensed only to trusted partners onboarded through the Foundry platform, meaning academia and competitors currently cannot independently assess these claims.
Why Compare Only with Open-Source Models?
An obvious question that few state outright: why does Microsoft list only open-weight models in its performance comparison, rather than comparing with closed-source models such as GPT-5, Gemini Ultra, or Claude Opus?
Discussion on Hacker News gives the most direct answer: because the weights and raw inference interfaces of top closed-source foundation models are not open to outsiders, external researchers simply cannot conduct foundation-model-level comparisons under equivalent conditions, and can only benchmark against available open-source models. This is a systemic predicament in foundation model evaluation across the entire industry, not Microsoft alone avoiding competition.
The open-source model lineup included in the comparison represents the current strongest level: DeepSeek-V4-Pro-Base (1.6 trillion total parameters, 49 billion active parameters), Kimi K3 (2.8 trillion total parameters, 104 billion active parameters), and Qwen3.8 (2.4 trillion total parameters, 95 billion active parameters). According to a specifications roundup by morphllm.com, DeepSeek-V4-Pro was released in April 2026 under the MIT license and supports a 1 million-token context window; Kimi K3 was released by Moonshot AI on July 27, 2026, and is the largest open-weight model to date.
MAI-Base-2 vs. Kimi K3: Similar Parameter Counts, Diverging Paths
In terms of parameter scale, MAI-Base-2 (2.8 trillion total parameters) is exactly the same as Kimi K3, but their architectural orientations differ markedly: MAI-Base-2 activates 80 billion parameters per token, while the corresponding figure for Kimi K3 is 104 billion, meaning the latter incurs heavier computation per inference, but the ceiling of its expressive capability also differs.
More noteworthy is the difference in positioning between the two. Kimi K3 is an open-weight model that can be downloaded and deployed immediately after release, and the community has already used it for fine-tuning and derivative work; MAI-Base-2's weights are currently licensed only to Foundry partners, and post-training (the instruction fine-tuning and reinforcement learning stage) has not yet been completed, with no release timetable for a final-user product form. In other words, MAI-Base-2 is still a "base model"—it is the starting point for all subsequent capabilities, not the endpoint.
Microsoft emphasizes the "organic" sourcing of its training data and does not include distillation from other models' outputs—a statement that carries special meaning in the current industry context. In recent years, using closed-source model outputs to train open-source models has sparked multiple legal and ethical controversies, and Microsoft's explicit statement here is both a technical claim and a compliance position. But according to explainx.ai, this point currently cannot be independently assessed.
MoE Architecture Has Become the New Default
MAI-Base-2, Kimi K3, and DeepSeek-V4-Pro, the three most representative large-scale foundation models of 2026, all adopt mixture-of-experts architecture. This convergence is no coincidence.
The core logic of MoE is to trade sparse activation for expansion of total parameter count. A 2.8-trillion-parameter MoE model activates only 80 billion to 104 billion of those parameters per inference, bringing actual per-token computation to less than one-tenth that of a dense model of comparable scale. This makes ultra-large-scale training engineering-feasible—Microsoft specifically mentions in its blog that by improving its "dropless multi-round routing (dropless routing)" scheme, it raised MoE model FLOPs utilization from 24% to 35%, a quite substantial advance at the systems engineering level.
In addition, MAI-Base-2 integrates visual capabilities directly into the pre-training stage rather than attaching them through a separate image encoder—an architectural choice that means visual understanding and language capabilities share representations at a foundational level, rather than being a patchwork of two systems. The 18% visual data share (about 4.5 trillion tokens) is relatively high among existing public technical reports.
Independent Judgment
The real significance of MAI-Base-2 is not how much stronger it is than competitors—that question must wait until post-training is complete and independent benchmark tests are available. Its significance lies in this: Microsoft is systematically internalizing foundation model research and development capability, no longer relying solely on external partnerships. From MAI-Base-1 to MAI-Base-2, and from the post-trained model MAI-Thinking-1 to the MAI-Base-3 already in preparation, this is a clearly visible path of capability accumulation.
The claim of a 4–12x efficiency improvement, if ultimately independently corroborated, would mean a substantial drop in the cost of training models of equivalent quality, with structural implications for the compute economics of the entire industry. But for now this remains an internal assertion. The real test will come after the post-training stage is complete: when a practically usable MAI-Thinking-2 or MAI-Chat-2 arrives and undergoes external evaluation on public benchmarks such as MMLU, MATH, and code generation, then the numbers will speak.
Sources: - [Frontier Pre-training at Microsoft AI](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/) - [Microsoft AI unveils 2.8 trillion-parameter model, claims efficiency gains](https://www.investing.com/news/stock-market-news/microsoft-ai-unveils-28-trillionparameter-model-claims-efficiency-gains-4952355) - [Microsoft completes pre-training of MAI-Base-2 AI foundation model](https://www.streetinsider.com/Corporate+News/Microsoft+completes+pre-training+of+MAI-Base-2+AI+foundation+model/27192270.html) - [DeepSeek V4 Pro: Model Overview, Features & Performance Guide](https://deepinfra.com/blog/deepseek-v4-pro-model-overview) - [Kimi K3 Model Overview: 2.8T Parameters, MXFP4 Quantization](https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei) - [Hacker News discussion: Frontier Pre-Training at Microsoft AI](https://news.ycombinator.com/item?id=)© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接