Microsoft's new MAI models
Simon Willison's Weblog (LLM / AI engineering) · Jun 2, 2026
Claim: Microsoft released two new models: MAI-Thinking-1 (1T parameters, 35B active, reasoning) outperforms Claude Sonnet 4.6 in blind evals; MAI-Code-1-Flash (137B parameters, 5B active) is purpose-built for GitHub Copilot with lower cost. Both trained on "clean and commercially licensed data" without third-party distillation.
Evidence: Microsoft official announcement; MAI-Code-1-Flash rollout to GitHub Copilot individual users in VS Code; MAI-Thinking-1 technical paper published; author attended Microsoft Build conference. Technical paper (p. 80+) reveals training data sourced from proprietary web crawl (1.2T pages, filtered to 794B) plus Common Crawl (24.2B pages post-filtering). No distillation claimed. Licensing: standard web crawl with filtering for adult/piracy content and AI-generated content, but fundamentally web-sourced, not exclusively "appropriately licensed".
Rationale
Real release (MAI-Code-1-Flash actively rolling out to Copilot users); MAI-Thinking-1 has published technical paper. Applicability to Timothy's CC+Codex+TS workflow is moderate: MAI-Code-1-Flash could be a cost/perf alternative for code tasks, but no evidence it integrates with CC or Codex. Claim of "appropriately licensed" data is undermined by the author's own update revealing standard web-crawl sourcing—this self-correction lowers confidence in the "clean data" narrative. MoE architecture (35B active of 1T) is interesting for cost but no token-spend data provided for Timothy's stack. Durability: depends on whether models remain competitive post next release. Commoditization_risk=0 because these ARE already commoditized products (released, in beta, on roadmaps). Source_credibility stays at 2 (credible outlet with direct conference access, but the original post contained major errors now corrected; the self-corrections preserve credibility but signal care is needed).