Microsoft finishes pre-training MAI-Base-2: 2.8T MoE, 80B active

Microsoft AI says it has completed pre-training of MAI-Base-2, its next foundation model after MAI-Base-1 / MAI-Thinking-1. Details are in Microsoft’s own technical post dated October 15, 2026.

Model size and training run

MAI-Base-2 is described as a mixture-of-experts (MoE) model with:

• 2.8 trillion total parameters
• 80 billion active parameters per token
• Native vision in pre-training (encoder-free VLM path), with vision data about 18% of a 50T-token mixture
• Trained on 16,000 GB200 GPUs for 38 days (August–September 2026); Microsoft reports average MFU 21% and goodput 89%

Microsoft stresses the run used organic data processed in-house (knowledge cutoff cited as March 2026 for web), not distillation from other labs’ models.

Claims vs open bases

On held-out bits-per-byte evaluations, Microsoft says MAI-Base-2 beat every open-weight model it tested, including DeepSeek-V4-Pro-Base. Fitting scaling laws, it estimates its recipe is 4.2–12.3× more pre-training FLOPs-efficient than that DeepSeek base. Those are Microsoft’s figures on Microsoft’s evals — not an independent benchmark suite.

Engineering notes include MoE in every layer after MoE MFU improvements, the Muon optimizer, and activation-gradient clipping for stability.

What’s next

MAI-Base-2 is in post-training. Microsoft says owning the stack should let it license weights to trusted partners through Foundry, but it has not published a consumer product date, API name or price. Work on MAI-Base-3 is already underway, per the same post.

For builders watching open and proprietary bases, MAI-Base-2 is Microsoft’s clearest statement yet that it wants a frontier base of its own — with the usual wait for post-trained checkpoints and third-party verification.