Microsoft AI has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters. It was built entirely in-house and is now in post-training. Microsoft plans to license the weights to trusted partners through Foundry, but there is no API, price or release date yet.
The practical point for anyone building on Azure is that “Azure means OpenAI” is no longer a safe long-term assumption. Make sure your model layer can be swapped.
What happened
According to Technology.org’s breakdown of the MAI-Base-2 training run, Microsoft trained the model on 16,000 NVIDIA GB200 GPUs over 38 days in August and September 2026. The training mix was 50 trillion tokens, with web data up to a March 2026 knowledge cutoff. Average model FLOPs utilisation was 21%, and the run made progress 89% of the time. For comparison, its predecessor MAI-Base-1 had 1 trillion total and 35 billion active parameters, and became the base for the MAI-Thinking-1 reasoning model in June.
The architectural change that matters most to builders is vision. MAI-Base-1 read only text. MAI-Base-2 learned images from the start of pre-training, using an encoder-free design with no separate vision module. Vision-language data made up 18% of the training mix and covered captions, charts, spatial reasoning, medical images and computer use.
OODA Loop’s brief on the announcement also lists untied embeddings, the Muon optimiser and MoE layers across the entire network. Microsoft stresses that the model learned from organic data with no distillation from other companies’ models. It presents that clean data trail as the basis for licensing the weights.
What the efficiency claim does and doesn’t say
The “4 to 12 times more efficient” headline needs careful reading. Microsoft says its recipe needs 4.2 to 12.3 times less compute than DeepSeek-V4-Pro-Base’s to reach the same quality. Quality here was measured as bits-per-byte on held-out tests built from its own internal codebase plus commissioned STEM, maths and QA sets.
As Technology.org points out, this compares training recipes, not finished models. Microsoft says it spent the saving on a larger model, not a cheaper one. Only DeepSeek released a base checkpoint, so the other rivals (Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL) were tested as post-trained models. No closed models appear in the published comparison.
The bolder claim came from outside the published report. Windows Mode, citing Business Insider, reports that Microsoft AI CEO Mustafa Suleyman told employees at a town hall that the model is stronger than the base model behind OpenAI’s GPT 5.6 Sol, while using a quarter of the compute. Nobody outside Microsoft can check that yet. Microsoft declined to comment to Business Insider. Treat it as a signal of intent, not a benchmark.
What it means for builders
Windows Mode notes that the biggest models in Foundry today come mostly from partners like OpenAI. A first-party, frontier-scale option changes the shape of the procurement conversation inside Azure.
Tech Insider argues that if Microsoft’s training efficiency translates into cheaper inference, Azure customers could see more competitive per-token pricing. That is a real if. Training efficiency and serving price are different numbers, and Microsoft has published neither a price nor serving costs.
Our read is that the quieter effect matters more than any price cut. Once Microsoft can credibly point to its own model, the leverage inside its OpenAI partnership shifts without either side making a public move. Technology.org frames the wider effort the same way: in-house models built to depend less on partners such as OpenAI, now paired with Microsoft’s own Maia 300 inference chip.
The provenance story matters too. Windows Mode suggests a first-party model with a known training data trail could appeal to regulated industries. If you are documenting model lineage for audit, our AI security and compliance guide for EU firms covers what that paperwork looks like in practice. Bear in mind that “trusted partners” suggests access to the weights will be limited at first.
What to do now
There is nothing to call yet, so this is about preparation, not migration. If MAI models arrive in the Foundry catalogue at a useful price, you want to be able to test them in days, not quarters.
- Put a thin gateway between your application and the model provider. Prompts, tool schemas and output parsing should not assume one vendor’s quirks. Our piece on integrating LLMs into legacy enterprise systems covers where that seam usually belongs.
- Build an eval set from your own traffic now. A few hundred real inputs with graded outputs tell you more than any bits-per-byte chart, and make a new model a measurable decision.
- Cost workloads per completed task, not per token. A cheaper model that needs more retries or a longer context can cost more overall. We break this down in how much it costs to build an AI agent.
- Watch the licence terms, not just the leaderboard. Weight licensing for approved partners is a very different proposition from a pay-as-you-go endpoint.
MAI-Base-2 is a large, well-documented training run with self-reported results and no product attached. The sensible response is not to bet on it. It is to make sure your architecture would not care which frontier model Azure sells you next year.
Frequently Asked Questions
Can I use MAI-Base-2 on Azure today?
No. The model is still in post-training. Microsoft has given no public API, release date or price. Its stated plan is to license the weights to trusted partners through Foundry.
Does 4-12x more efficient mean MAI-Base-2 will be 4-12x cheaper to use?
Not necessarily. Microsoft's 4.2 to 12.3 times figure compares pre-training recipes against DeepSeek-V4-Pro-Base. Microsoft says it spent that saving on training a larger model. Inference pricing has not been announced.
Has MAI-Base-2 been shown to beat OpenAI's models?
Not publicly. Microsoft's published comparisons cover open-weight models only. The claim that it beats the base model behind GPT 5.6 Sol came from Mustafa Suleyman at an internal town hall, as reported by Business Insider, and has not been independently verified.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Set up a free 30-minute call and we will map your first win together.
Get in touch
