On October 1, 2026, Arena officially announced that Xiaomi's MiMo-V2.6-Pro and MiMo-V2.6-Flash have landed on the Agent Arena leaderboard. This is the first large-scale test in real agent sessions for the MiMo-V2.6 series since its September 21 release and subsequent open-sourcing under the MIT license.
Start with Pro's report card: across 8,100+ real agent sessions, Pro posted a +3.17% net improvement, ranking 5th among open-source models. The comparison is telling — the previous generation, MiMo-V2.5-Pro, sits at 13th on the same leaderboard with a -7.23% net improvement. In one generation: up 9 places and 10.4 percentage points. Even more worth watching is the Confirmed Success signal: Pro scored +7.35% there, 2nd among open-source models. Flash plays the value game: 9th among open-source models with a -0.57% net improvement, but a median task cost of just $0.04 — 56% cheaper than Pro — also landing on Arena's cost-performance frontier.
Confirmed Success deserves its own explanation. Agent Arena doesn't test with multiple choice; models do real work on real tasks, and Confirmed Success measures the quality signal of tasks confirmed complete — stricter than merely "it ran." Pro ranking 2nd among open-source models on this signal means it isn't just farming session volume; it's genuinely more reliable on tasks requiring multi-step decisions. We previously covered the open-source release of MiMo-V2.6, when it topped the open-source charts with 46 points; this Agent Arena result fills in the "real task execution" piece of the puzzle.
One caveat about leaderboards: Arena scores come from votes on real user sessions and shift with sample size and task distribution — +3.17% is net improvement over a baseline, not an absolute win rate. Reading it as evidence that "each generation is stronger" holds up; reading it as "already unbeatable" goes too far.
For the open-source model landscape, the signal is clear: competition in agent scenarios has moved from "parameters and leaderboard scores" to "reliable delivery on real tasks." MiMo-V2.6-Pro took one generation to flip net improvement from negative to positive and pushed Confirmed Success into the open-source top two, suggesting Xiaomi's scaled reinforcement-learning approach is paying off in agent workflows. For users, the choice gets clearer too: Pro for the ceiling, Flash for the budget — performance and price tiers are now cleanly separated within the same series.
