Microsoft Finishes Pre-Training MAI-Base-2, a 2.8-Trillion-Parameter MoE It Says Needs 4 to 12 Times Less Compute Than DeepSeek -- and Trained Without Distilling Anyone Else's Model

On October 15, 2026, Microsoft AI published "Frontier Pre-training at Microsoft AI", announcing that it has "recently completed pre-training our new state-of-the-art model, MAI-Base-2, an MoE foundation model with 80 billion active and 2.8 trillion total parameters." It is the base model behind Microsoft's next generation of in-house models, and it "is currently undergoing post-training".

The training run

• Hardware and time: "16k GB200 GPUs over 38 days in August and September 2026".
• Efficiency of the run: average model FLOPs utilisation of 21% and goodput of 89%.
• Data: a 50T-token mixture with a March 2026 knowledge cutoff. Investing.com reports the unique-token corpus grew about 36% over MAI-Base-1, which Microsoft's post confirms.
• Vision: unlike MAI-Base-1, it is "pre-trained natively with vision using an encoder-free VLM architecture".

How the efficiency claim is built

The headline is that MAI-Base-2 "outperforms every open-weight model we tested on our held-out evaluations" and "is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base". The method matters for reading that number. Models are compared using bits-per-byte (BPB) on held-out text, code from Microsoft's own codebase, and STEM, math and Q&A sets "commissioned from vendors to reduce the risk of training-set contamination". For each evaluation Microsoft fits a scaling law of BPB against training FLOPs and estimates what compute its recipe would need to match DeepSeek, which gives "4.2–12.3x". Microsoft also notes that Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL are available only as post-trained checkpoints, "with the caveat that post-training may affect these evaluations."

Why organic data is a business point

The post states: "Our pre-training is built on organic data which does not involve distillation from other models. This will enable us to license our model weights to trusted partners through Foundry". Microsoft is signalling to enterprises that a licensed MAI model will not carry provenance questions about outputs taken from rival labs, a live issue across the industry this year. Investing.com reports Microsoft gave no availability dates or pricing. Preparations for MAI-Base-3 "are already underway".

What remains uncertain. Everything here describes a base model. Bits-per-byte predicts downstream quality only loosely, and the efficiency multiple is an estimate from fitted curves, not a head-to-head training run. How the post-trained model performs on public tasks, and on what terms partners can license the weights, is still unknown.

Analysis: Microsoft is building a second supply line next to its OpenAI partnership, and this post is a pitch to the customers who would use it. For engineering teams, the useful details are less the parameter count than the method. Scaling ladders, hyperparameter laws fitted at small scale, and vendor-commissioned, decontaminated evaluations are a template for making training decisions before spending the compute. The no-distillation claim is worth noting for procurement teams too. Provenance of training data is becoming a feature that model vendors compete on, and buyers can ask every vendor the same question.