Xiaomi's MiMo Pro jumps from 58.41 to 72.57 on DeepSWE after RL runs
nrehiew_ · x · 2026-09-21
Xiaomi's MiMo RL training runs have completed, with the MiMo Pro model improving from 58.41 to 72.57 on the DeepSWE benchmark. For context, the current DeepSWE leaderboard top is 74, held by Astra, Gemini 3.8 Flash and Opus 5, putting MiMo Pro close to the leading tier.
More from Models
- Grok bot is now fully integrated inside the Grok app — ns123abc · 2026-09-21
- Judge Models Are Saturating Benchmarks Fast, But Who Owns Your Option Table? — sujingshen · 2026-09-21
- GLM 5.3 Regresses: Users Find It Worse Than GLM 5.2 on Arena — floriandotorg · 2026-09-21
- Cua Open-Sources Jev-Style 706K-Param Form Model, Beats Jev API 99.7% vs 83.6% — aigclink · 2026-09-21
- Jev as a Model Router Might Be Its Killer Use Case — nicolechirps · 2026-09-21
- Industry consensus: agent harnesses will specialize per task, moat lives in the data layer — amankhan · 2026-09-21