Microsoft releases MAI Base 2 with superior pre-training efficiency
Microsoft announced MAI Base 2, a 2.8 trillion parameter model trained on 16,000 GPUs in 38 days. The company claims the model outperforms open competitors and uses a fraction of the compute required for rival frontier models.
Microsoft AI CEO Mustafa Suleiman announced the release of MAI Base 2, a 2.8 trillion parameter base model utilizing a mixture of experts architecture with 80 billion active parameters. The model underwent a fresh pre-training run on 16,000 GB200 Blackwell GPUs over 38 days, with no distillation involved. Suleiman stated internally that MAI Base 2 is stronger than the base model powering GPT 5.6 Sol and required only one-quarter of the compute used by OpenAI for that model, a comparison derived from proprietary data shared under Microsoft's investment deal.
Publicly, Microsoft asserts that MAI Base 2 outperformed all open models exceeding one trillion parameters on pre-training benchmarks, including Kimi K3. The company also reported that the model achieved more than 12 times the pre-training efficiency of DeepSeek v4 Pro. Although MAI Base 2 is currently a base model without reinforcement learning, Microsoft is deploying it alongside its Frontier Tuning business model. This approach allows enterprises to fine-tune their own models on Microsoft bases, addressing cost and sovereignty concerns. Microsoft argues that owning the intelligence layer through fine-tuning is imperative for enterprises, offering a sovereignty-focused alternative to relying solely on OpenAI or Anthropic.