MiroThinker-1.7 tops BrowseComp with verification agents

MiroThinker is making the case that scaling *interaction depth* — not just parameters or context — is the missing axis for agents that do real research, not just plausible retrieval.

• –The verification-centric architecture is the real differentiator: local verification breaks probability bias at each reasoning step, global verification audits the full evidence chain end-to-end — and paradoxically, verified runs use *fewer* steps than unverified ones by filtering no-info-gain actions
• –BrowseComp 88.2% (H1) is a meaningful signal — this benchmark tests genuine web research ability, not memorization, making it harder to game than static evals
• –The 30B mini outperforming Kimi-K2-Thinking (1T parameters) on BrowseComp-ZH at roughly 1/20th the inference cost ($0.07 vs $1.40 per call) is a striking efficiency claim worth watching
• –Full open-source release includes weights, training code, and the MiroVerse 147K-sample dataset — unusual transparency for a frontier research agent
• –6,700+ GitHub stars suggests strong developer traction despite flying under mainstream radar