started · updated
Microsoft unveils MAI-Base-2 foundation model with 2T parameters
Microsoft has completed pre-training for MAI-Base-2, a next-generation foundation model featuring 2 trillion parameters. The model utilizes a Mixture-of-Experts (MoE) architecture, activating approximately 80 billion parameters during inference to enhance computational efficiency.
In internal benchmarks covering science, programming, and mathematics, Microsoft reports that MAI-Base-2 outperformed leading open-source alternatives, such as DeepSeek-V4-Pro-Base. The company claims the model is up to 12 times more hardware-efficient than its competitors. The training process lasted 38 days, powered by 16,000 Nvidia GB200 chips, using a dataset of 50 trillion tokens.
MAI-Base-2 functions as a vision-language model (VLM), with approximately one-fifth of the training tokens dedicated to an integrated computer vision module. This allows for advanced tasks such as analyzing scientific charts and spatial reasoning. Microsoft plans to provide model weights to enterprise customers through its Microsoft Foundry platform and has already begun preparations for the subsequent MAI-Base-3 model.
Entities
DeepSeek · MAI-Base-2 · Microsoft · Nvidia
Claims
What the coverage asserts, and how many sources carry each claim.
- [○ 1 SOURCE] Microsoft has started development on the next model, MAI-Base-3. zdnet.co.kr
- [● 3 SOURCES] The model uses a Mixture-of-Experts (MoE) architecture that activates 80 billion parameters per prompt. www.prodiris.fr · siliconangle.com · zdnet.co.kr
- [● 2 SOURCES] The training dataset consisted of 50 trillion tokens. www.prodiris.fr · siliconangle.com
- [● 3 SOURCES] Training took 38 days using 16,000 Nvidia GB200 chips. www.prodiris.fr · siliconangle.com · zdnet.co.kr
- [● 3 SOURCES] The model outperformed DeepSeek-V4-Pro-Base on internal science, programming, and math benchmarks. www.prodiris.fr · siliconangle.com · zdnet.co.kr
- [● 2 SOURCES] MAI-Base-2 is up to 12 times more hardware-efficient than leading open-source alternatives. siliconangle.com · zdnet.co.kr
- [● 2 SOURCES] Approximately one-fifth of the tokens were used to train an integrated computer vision module. www.prodiris.fr · siliconangle.com
- [● 3 SOURCES] Microsoft developed the MAI-Base-2 foundation model with 2 trillion parameters. www.prodiris.fr · siliconangle.com · zdnet.co.kr