MAI-Code-1
#11 in Coding Agentsmicrosoft · v1 · seit 2026-06-02 · 30× · zuletzt 29. Juni 2026
36
Momentum
MAI-Code-1-Flash is Microsoft's first in-house coding model, unveiled at Build 2026 on June 2, 2026. It is a sparse Mixture-of-Experts (MoE) model with 137 billion total parameters and 5 billion active parameters per inference forward pass, with a 256,000-token context window. The model was trained specifically against GitHub Copilot's production harness and Visual Studio Code, targeting low latency, high token efficiency, and everyday developer workflows rather than frontier-level generality. It is integrated into GitHub Copilot (Free, Student, Pro, Pro+, Max, Business, Enterprise) and accessible via the model picker in VS Code and other Copilot surfaces.
Momentum trend
04.04.03.07.
Features
| License | Proprietary License (restricts commercial reuse outside Microsoft/GitHub products) |
| Price | $0.75 / 1M input tokens · $0.075 / 1M cached input tokens · $4.50 / 1M output tokens (GitHub Copilot usage-based billing / AI credits) |
| Release Date | June 2, 2026 (rollout start for GitHub Copilot Individual in VS Code; GA for Business/Enterprise: June 26, 2026) |
| Interface (IDE/CLI/Web) | VS Code (model picker, inline chat, agentic coding); GitHub Copilot CLI (auto model selection); GitHub.com Chat; GitHub Mobile; Visual Studio; JetBrains IDEs; Eclipse; Xcode; Copilot Cloud Agent |
| SWE-bench Verified (%) | 71.6% (vendor-reported, measured with GitHub Copilot production harness; for comparison: Claude Haiku 4.5 achieves 66.6%) |
