Microsoft AI Develops MAI Code 1.1 Flash Mixture-of-Experts Model
Microsoft AI has introduced MAI Code 1.1 Flash, a coding-optimized mixture-of-experts model featuring 137 billion total parameters and 6.8 billion active parameters. This model is part of a local development initiative announced alongside sandboxed tools for Windows. This release provides developers with a powerful, specialized local AI model that balances a massive knowledge base with efficient inference speeds. It highlights the growing trend of deploying advanced coding assistants locally rather than relying exclusively on cloud services. The model uses a Mixture-of-Experts architecture where a large total parameter count of 137B allows for broad knowledge retention, while only 6.8B active parameters are used per token to ensure fast execution. It is tailored specifically for code generation and development tasks.
## BACKGROUND
A mixture-of-experts (MoE) is a neural network architecture that activates only a subset of its total parameters for any given input, achieving high capacity without the heavy computational cost of dense models. Total parameters represent the entire model capacity across all expert networks, while active parameters refer to the subset engaged during a single token generation.