MAI Code 1.1 Flash
A programming-optimized mixture-of-experts (MoE) code model released by Microsoft AI, with 137 billion total parameters and 6.8 billion active parameters, supporting local deployment and sandboxed tool calling.
Timeline (last 90 days)
Microsoft AI developed a local version of MAI Code 1.1 Flash, a Mixture-of-Experts (MoE) model optimized for programming.
The local version of MAI Code 1.1 Flash has a total parameter count of 137 billion, activating only 6.8 billion parameters per inference.
Even with quantization, the VRAM usage of a 137B model still far exceeds the VRAM limit of a single consumer-grade GPU, and multi-GPU splitting or CPU+GPU hybrid inference is often a necessary option.
Running the model locally can eliminate compliance concerns about uploading code to cloud services, and can be combined with sandboxed tools to complete a closed loop of code generation, testing, and debugging.
Reddit community members pointed out that even with quantization, the storage and loading costs of deploying 137B total parameters are still not low, and the official side has not yet given clear hardware requirements.
All Facts (5)
The local version of MAI Code 1.1 Flash has a total parameter count of 137 billion, activating only 6.8 billion parameters per inference.
50%UnverifiedEven with quantization, the VRAM usage of a 137B model still far exceeds the VRAM limit of a single consumer-grade GPU, and multi-GPU splitting or CPU+GPU hybrid inference is often a necessary option.
50%UnverifiedRunning the model locally can eliminate compliance concerns about uploading code to cloud services, and can be combined with sandboxed tools to complete a closed loop of code generation, testing, and debugging.
50%UnverifiedReddit community members pointed out that even with quantization, the storage and loading costs of deploying 137B total parameters are still not low, and the official side has not yet given clear hardware requirements.
50%UnverifiedMicrosoft AI developed a local version of MAI Code 1.1 Flash, a Mixture-of-Experts (MoE) model optimized for programming.
50%