Key Findings
Microsoft has completed the pre-training of MAI-Base-2, a next-generation foundation model featuring 2.8 trillion total parameters. Internal evaluations indicate that the model reduces the compute required to achieve performance equivalent to DeepSeek-V4-Pro-Base by 4.2x to 12.3x. Additionally, training efficiency has improved 6.3x to 12.7x compared to the previous MAI-Base-1 model.
Technical Details
The model employs a Mixture-of-Experts (MoE) architecture, where only approximately 80 billion parameters are activated during inference. Training took 38 days between August and September, utilizing 16,000 Nvidia GB200 GPUs and 50 trillion tokens, 18% of which were visual data. Microsoft improved model compute utilization in the MoE segment from 24% to 35% by optimizing memory management and applying the “Muon” optimization algorithm alongside gradient adjustment techniques.
MAI-Base-2 is a native vision-language model (VLM) that trains on text and images simultaneously without a separate image encoder. This allows it to perform spatial reasoning, computer use, and document analysis. Microsoft confirmed that the model was trained using proprietary source data rather than distillation from other AI outputs, increasing unique training data by 36% over the previous version.
Background & Context
Parallel to model development, AI-driven software rewriting is showing significant results. Developer Theo Browne used AI agents to port TypeScript development tools to Rust (TS_Rust), achieving speeds 12.5x faster than existing TypeScript implementations and 2x faster than Microsoft’s Go-based version. Browne noted that switching from OpenAI models (costing over $400,000 with 84% compatibility) to Anthropic’s Claude (costing approximately $24,000) allowed him to pass most official compatibility tests.
Significance & Outlook
Microsoft plans to offer MAI-Base-2 weights to enterprise customers via “Microsoft Foundry” and has already begun preparations for MAI-Base-3. The industry trend toward Rust—seen in Microsoft’s Copilot runtime rewrite and Google’s use of Gemini 4 Argon for C/C++ porting—suggests a move toward lighter, faster runtime environments that could lower operational costs for AI coding services by eliminating garbage collection and improving WebAssembly compilation.
Source: https://finance.biggo.com/news/628565eb-7cc6-46b5-a61f-8be6fcbebc5b
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
