The MAI-Base-2 release date is still a blank. Microsoft AI said on October 15 that it finished pre-training this 2.8-trillion-parameter model, but you can’t download it, call it through an API, or buy it yet.
Think of it as the raw engine, not the finished car. Here’s what Microsoft actually announced, and what it didn’t.
Quick facts, so you can skim:
- 2.8 trillion total parameters, with 80 billion active at a time.
- Pre-training is finished. Post-training is still running.
- No public weights, API, price, or release date.
- Microsoft plans to license weights to “trusted partners” through Foundry.
That is the whole public picture today. The sections below explain what it means and what to watch.
What Microsoft says MAI-Base-2 is
MAI-Base-2 is a mixture-of-experts model. Only a slice of its parameters, about 80 billion, works on any one token, which keeps running costs lower than a dense model of the same size.
Microsoft says it trained on a 50-trillion-token mix over 38 days in August and September. The model handles images natively, and its knowledge cutoff is March 2026.
These are the main specs from Microsoft’s own Lab Notes post:
| Spec | MAI-Base-2 |
|---|---|
| Total parameters | 2.8 trillion (mixture-of-experts) |
| Active per token | 80 billion |
| Training data | 50 trillion tokens, 18% vision-language |
| Hardware | 16,000 GB200 GPUs, 38 days |
| Status | Pre-training done, post-training under way |
Microsoft also says the data is organic and wasn’t distilled from other models. That’s a company claim, and nobody outside Microsoft has checked it.
The 4-12x efficiency claim, explained
Microsoft says MAI-Base-2 is 4 to 12 times more efficient to pre-train than DeepSeek-V4-Pro-Base, which it calls the strongest open-weight base model. In plain terms, it claims to reach similar quality with far less compute.
It also says the model beats every open-weight model it tested on its own held-out evaluations. Some of those tests are Microsoft’s own, and others were commissioned from vendors.
So treat the numbers as a strong hint, not proof. A base model is also not what you chat with. Independent benchmarks only mean something once Microsoft ships the post-trained version.
MAI-Base-2 release date: why you can’t use it yet
A base model has only learned to predict text. Post-training is the stage that turns it into something that follows instructions and holds a conversation, and Microsoft says that stage is under way.
Microsoft’s earlier MAI-Thinking-1, announced in June, was built the same way on top of MAI-Base-1. Expect a similar path here: a named, post-trained model first, then access.
Here’s what is and isn’t known right now:
| Question | Answer today |
|---|---|
| Can I download it? | No. No public weights. |
| Is there an API? | No endpoint has been announced. |
| What does it cost? | No price published. |
| When does it ship? | No date. Microsoft says only that more results are coming soon. |
| Who gets it first? | “Trusted partners” through Foundry, per Microsoft’s stated plan. |
If you want a sense of how Microsoft prices and ships models in Foundry, our look at Microsoft Decision-1 is a useful reference point.
Where this fits in Microsoft’s AI plans
Microsoft says its approach lets it tune models first in its own products and then for Foundry customers. That points to Copilot-style features getting upgrades before outside developers do.
It also fits the hardware story. Our guide to the Maia 300 AI chip covers the silicon Microsoft wants to run these models on from 2027.
Microsoft adds that work on MAI-Base-3 has already started, with no timeline. The message is clear: it wants its own frontier models, not only partner ones.
What to watch before you get excited
Three signals will tell you when this model becomes real. Each one is something Microsoft hasn’t announced yet.
- A product name for the post-trained model, since “MAI-Base-2” is the raw base.
- A Foundry listing with a price per million tokens.
- Third-party benchmarks from people who don’t work at Microsoft.
Until at least one of those lands, the model is a research milestone. It shows Microsoft can train at the 2.8-trillion scale on its own, which matters for its long-term independence from partners, but it isn’t something you can use.
Competing labs are shipping open weights right now. Meta’s Muse Spark Open is already downloadable, which makes Microsoft’s closed, partner-only plan look cautious by comparison.
Frequently Asked Questions
When will MAI-Base-2 be released?
Microsoft hasn’t said. It only says post-training is under way and that it will share more results soon.
Is MAI-Base-2 open source?
No. Microsoft plans to license the weights to trusted partners through Foundry, not publish them for everyone.
How big is MAI-Base-2?
It has 2.8 trillion total parameters, with 80 billion active per token. That makes it one of the largest base models announced so far.
Can I try MAI-Base-2 in Copilot today?
Not as far as Microsoft has said. No product has been named as running on it yet.
My take: nothing here changes your day yet, so don’t plan around it. Watch for the post-trained model’s name, a Foundry listing, and independent benchmarks. Until those show up, the 4-12x claim is Microsoft grading its own homework.
Leave a Reply