Microsoft finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model
Microsoft AI, Thursday: MAI-Base-2, a mixture-of-experts model with 80 billion active parameters and 2.8 trillion in total, trained on 16,000 GB200 GPUs over 38 days in August and September. The lab says the recipe is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. Vision is 18 percent of a 50-trillion-token mix. No release date or price. Post-training is underway, and work toward MAI-Base-3 has started.
Deep dive3 minIn-depth
Maya Chen3 min read
- Microsoft
- MAI
- Pre-training
- Foundry

Photo: Microsoft AISource page
Microsoft AI said on Thursday that it has finished pre-training MAI-Base-2, a mixture-of-experts foundation model with 80 billion active parameters and 2.8 trillion in total. In a lab note dated 15 October, the company said the model beats every open-weight system it tested on held-out evaluations, and that the training recipe is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which it called the strongest open base model. MAI-Base-2 is in post-training. The note gives no release date and no price. Preparations for MAI-Base-3, it said, are already underway.
The run used 16,000 GB200 GPUs over 38 days in August and September 2026. Microsoft said average model FLOPs utilization was 21 percent and goodput — the fraction of time training progressed — was 89 percent. It normalizes that utilization figure by 2.5 petaflops, which it calls the GB200’s theoretical maximum in BF16, and excludes FLOPs for rematerialization and the optimizer. The training mixture was 50 trillion tokens. Unlike MAI-Base-1, MAI-Base-2 was pre-trained with vision, using an encoder-free vision-language architecture. Microsoft said its vision-language corpus holds 4.5 trillion unique tokens and, with repetition, accounted for 18 percent of the mix. The tasks it listed include image captioning, chart understanding, visual grounding, spatial reasoning, medical understanding, computer use and sheet-music reading. Unique training tokens grew about 36 percent from MAI-Base-1. Web data, Microsoft wrote, has a March 2026 cutoff.
What the efficiency figure measures
The 4.2 to 12.3 times claim is an estimate of the recipe, not a public scoreboard of finished chatbots. Microsoft said it fit a scaling law of bits-per-byte against training FLOPs for each of four held-out evaluations, estimated how much compute a model on its own ladder would need to match DeepSeek-V4-Pro-Base, and compared that estimate with DeepSeek’s training compute. It included MAI-Base-2’s final result in the fits, so the matching point falls inside the range of models it trained. DeepSeek-V4-Pro-Base, the note says, is the only released base checkpoint above one trillion parameters, and it was still trained on significantly fewer FLOPs than MAI-Base-2. Against its own June model, MAI-Base-1, Microsoft put the FLOPs-efficiency gain at 6.3 to 12.7 times. It attributed the gain to data and to modeling choices.
The other names on the chart are not base models. Microsoft compared MAI-Base-2 with frontier checkpoints available before October 2026 and listed Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL alongside DeepSeek-V4-Pro-Base. Only post-trained checkpoints are available for those three, it wrote, and post-training may affect the scores. No closed model is in the comparison. The tests are Microsoft’s. Held-out code is the lab’s own codebase. STEM, math and question-answering sets were commissioned from vendors, Microsoft said, to reduce the chance of training-set contamination, and every evaluation set was decontaminated against the training data. Bits-per-byte is used so the score does not depend on a tokenizer. Lower is better. On that measure, Microsoft said MAI-Base-2 came out ahead of the models it tested.
The data section is also the commercial argument. Microsoft said pre-training is built on organic data, processed in-house, and does not involve distillation from other models. That provenance, it wrote, “will enable us to license our model weights to trusted partners through Foundry.” The weights are not described as a public download. Owning the stack, the note says, lets Microsoft tune models and infrastructure together, first in its own products and then for Foundry customers. Investing.com, in a story published at 12:26 p.m. on Thursday, reported the same parameter counts, the 16,000-GPU run, the 4.2 to 12.3 times range and the 18 percent vision share, and said Microsoft did not disclose commercial availability or pricing. StreetInsider’s item, timestamped 12:18 p.m. Eastern the same day, carried the completion of pre-training and the move into post-training.


