Microsoft Ships MAI-Cyber-1-Flash — 137B Total / 5B Active MoE That Pushes MDASH From 88.4% to 95.95% on CyberGym at Half the Cost

Microsoft AI (corroborated by Axios, VentureBeat, The Hacker News, The Register, MarkTechPost)·high signalXBlueskyLinkedInCopy link

Microsoft AI released MAI-Cyber-1-Flash on 2026-07-27, a sparse-MoE security model with 137B total parameters, 5B active, and a 256K context window — a cybersecurity fine-tune of MAI-Code-1-Flash, the same lightweight agentic coder already inside GitHub Copilot and VS Code. Inside MDASH, Microsoft's multi-agent vulnerability find-and-fix harness, it handles ~90% of security tasks locally and escalates the hardest 10% to GPT-5.4, lifting CyberGym from 88.4% to 95.95% while costing 50% less than the prior GPT-5.4 + 5.4-mini + 5.3-Codex configuration. Microsoft claims it tops GPT-5.5 Cyber, Mythos 5, GPT-5.6 Sol and Gemini 3.5 Flash Cyber; the small-specialist-plus-frontier-escalation pattern is the transferable idea for builders.