Skip to content
Microsoft AIPreview

MAI-Code-1-Flash

Tiny (5B), IDE-native coding model — the per-keystroke lane, not the architect.

Read the full MAI-Code-1-Flash analysis

Context

—

Max output

—

Input /1M

—

Output /1M

—

Best for

  • In-editor completion in VS Code / GitHub Copilot CLI
  • High-frequency, low-latency coding assists

Watch out

Vendor-claimed 51% SWE-Bench Pro. Benchmark size ≠ how it feels on your repo — pilot before adopting.

For creators. Cheap, fast in-editor help; route real architecture work to a frontier model.

Benchmarks

swe bench pro51

Capabilities

  • Inference-efficient coding model (~5B params, Haiku-class size)
  • Tuned for VS Code and GitHub Copilot CLI
  • In-editor, high-frequency / low-latency use

More from Microsoft AI

Microsoft Phi-4 (open-weight family)MAI-Thinking-1MAI-Image-2.5

Sources