MAI-Base-2 is not a public download. Microsoft’s pre-training note, published 15 October 2026, says pre-training of MAI-Base-2 is complete. It is a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters. The same note says the model is currently undergoing post-training, and that licensing the weights to trusted partners through Foundry is future work (“this will enable us”). An API price, a context length, a quantization, and a local VRAM number are not published.
Figures Microsoft published
| Item | Value | What it is |
|---|---|---|
| Active parameters | 80 billion | Model scale in the note |
| Total parameters | 2.8 trillion | Model scale in the note |
| Training cluster | 16k GB200 GPUs | Training, not inference |
| Training time | 38 days, August and September 2026 | Training |
| Average MFU | 21% | Normalized by 2.5 PFLOPS, GB200 BF16 theoretical max, excluding rematerialization and optimizer FLOPs |
| Goodput | 89% | Fraction of time training progressed |
Microsoft also says MAI-Base-2 outperformed every open-weight model they tested on their held-out evaluations, and that it is 4–12× more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. Those are Microsoft’s comparisons. This page does not turn them into a leaderboard or an inference-speed claim.
The note says the training run does not involve distillation from other models. Secondary write-ups of token counts or layer counts are not repeated here, because they were not in the sentences checked on the Microsoft page.
Sources: Frontier pre-training at Microsoft AI and the linked PDF. Checked 16 October 2026.
Local run and price
No official Hugging Face model id for MAI-Base-2 was found on 16 October 2026. Without a checkpoint, file bytes and VRAM are not published. There is no calculator preset.
Foundry partner licensing is described in the future tense. This page does not treat that sentence as a current public weight release or as a price.
Open-weight models with files this site can describe are on the model directory and in the 16 Oct 2026 notes, including Step 5 Preview and Muse Spark Open. Calculated fits stay on catalog models such as Qwen3.8-27B. The formula limits are in the methodology.
FAQ
Can I download MAI-Base-2 and run it locally?
Not as a public checkpoint on 16 Oct 2026. Microsoft says the weights will be licensed to trusted partners through Foundry. The same post says MAI-Base-2 is still undergoing post-training. No Hugging Face repo, file size, quantization, or VRAM number was published, so this site does not estimate one.
How many parameters does MAI-Base-2 have?
Microsoft's 15 Oct 2026 pre-training note says MAI-Base-2 is a mixture-of-experts foundation model with 80 billion active parameters and 2.8 trillion total parameters. No layer widths, precision, or checkpoint byte size are in that note.
What hardware trained MAI-Base-2?
Microsoft says it trained MAI-Base-2 on 16k GB200 GPUs over 38 days in August and September 2026. Average MFU was 21 percent and goodput was 89 percent. MFU is normalized by 2.5 PFLOPS, which Microsoft calls GB200's theoretical maximum in BF16. That is a training report, not an inference memory budget.
Is there an MAI-Base-2 API price?
Not published. The note does not list a model id, a per-token price, or a context length.