Microsoft completes MAI Base 2 pretraining while assistant training continues
Microsoft has completed pretraining MAI Base 2, a mixture of experts model with 2.8 trillion parameters and 80 billion active. Its October 15 research update says posttraining is underway. This is not a finished assistant launch.
The company reports a 38 day run on 16,000 GB200 GPUs. The model was trained with vision from the outset, using a 50 trillion token mixture.
What the efficiency claim measures
Microsoft estimates that its recipe needs 4.2 to 12.3 times fewer training FLOPs to match DeepSeek V4 Pro Base on its evaluations. That comparison uses fitted scaling curves and bits per byte on held out material. It is not a measured reduction in the serving price of a finished assistant. Microsoft also cautions that some comparison checkpoints had already undergone posttraining.
A training efficiency result answers a narrower question than a product comparison. An application team still needs to establish whether a model follows its instructions, completes its tasks and meets its response time budget. A lower training loss alone does not settle those questions. Our coverage of GPU efficiency research at Flapping Airplanes examines the related distinction between a technical efficiency claim and evidence of savings on a deployed workload.
A base model is an intermediate result
Microsoft’s training overview distinguishes learning to predict text from learning to behave as an assistant. Pretraining produces the base model. Later stages can teach instruction following through curated examples and improve behavior using feedback or rewards.
That distinction is especially important for agents. Producing a valid tool request is learned behavior, while external software performs the requested action. A strong text predictor and a dependable agent are not interchangeable descriptions. The surrounding application must still execute tools and handle their results.
The earlier model shows the product gap
For comparison, Microsoft’s MAI Thinking 1 announcement, updated August 12, describes a reasoning model in public preview through Foundry. It lists a 256,000 token context window, function calling and compatibility with the Chat Completions API. Those are concrete integration details for a model developers can evaluate, not specifications to transfer automatically to a future model derived from MAI Base 2.
Microsoft says MAI Base 2 will underpin its next generation of models and discusses licensing weights to trusted partners through Foundry. The update does not announce a public download or API price.
For developers, the next test is practical. Access terms, task completion and response time need to be assessed on their own workloads.
Archival Microsoft campus photograph from June 2015 by Jiaqian AirplaneFan under CC BY 3.0. Converted to WebP. This does not depict the model or its training infrastructure.
