One-line summary
Microsoft AI announced on October 15, 2026 that it finished pre-training its in-house foundation model MAI-Base-2: MoE architecture, 2.8 trillion total parameters and 80 billion active per token, trained on 16,000 NVIDIA GB200 GPUs over 38 days (about 50 trillion tokens); as of the date of this dataset the model is still in post-training.
Data Notes
This dataset records the key parameters of MAI-Base-2 disclosed in Microsoft AI's technical announcement of October 15, 2026 (cross-checked via IT Home and AICoder reports). Data is current as of October 16, 2026; model weights, pricing and the commercial release date had not been announced as of collection.
Core Data
| field | meaning | value |
|---|---|---|
| model_name | Model name | MAI-Base-2 |
| architecture | Architecture | Mixture-of-experts (MoE): 1 shared + 7-of-512 routed experts per layer |
| total_params | Total parameters | 2.8 trillion |
| active_params_per_token | Active parameters per token | 80 billion |
| layers | Number of layers | 102 |
| vision | Vision capability | Native encoder-free vision |
| training_gpus | Training GPUs | 16,000 NVIDIA GB200 |
| training_days | Training duration | 38 days (August–September 2026) |
| training_tokens | Training data volume | About 50 trillion tokens |
| knowledge_cutoff | Knowledge cutoff | March 2026 |
| efficiency_claim | Compute efficiency (Microsoft claim) | 4.2–12.3x more compute-efficient than DeepSeek-V4-Pro-Base |
| performance_claim | Performance claim (Microsoft claim) | Outperformed every open-weight model evaluated |
| status | Current status | Pre-training complete, post-training in progress; weights, pricing and release date not announced |
| announce_date | Announcement date | 2026-10-15 (Microsoft AI official blog) |
Field Notes
- architecture: each layer has 1 shared expert plus 7 experts selected from 512 routed experts — a deep MoE design.
- active_params_per_token: the parameter scale actually activated per forward pass, directly affecting inference cost.
- efficiency_claim: Microsoft says MAI-Base-2 pre-training is 4.2–12.3x more compute-efficient than DeepSeek-V4-Pro-Base (sources differ slightly; this dataset uses Microsoft's official figure and labels it as a Microsoft claim).
- status: the model is in post-training and not yet available; the official release date follows Microsoft's announcement.
Data Sources & References
- IT Home: Microsoft completes pre-training of SOTA base model MAI-Base-2 (2026-10-16): link
- AICoder: Microsoft AI finishes pre-training MAI-Base-2 (2026-10-16): link
- Microsoft Foundry Labs (official): Innovations