Reasoning Models Need a New Measurement: Public Leaderboard MERA Reason Unveiled

Developers have launched a public leaderboard, MERA Reason, for evaluating reasoning models in Russian, consisting of four task sets. The first set, Luzitania, includes 251 complex Olympiad-level mathematical problems, filtered by difficulty using the GPT-OSS-120B model; the problems have been translated into Russian and require long chains of reasoning. The second set, TMath, contains 310 problems of varying difficulty from Russian Olympiads (All-Russian and Moscow), focusing on combinatorics, number theory, and geometry; the median reasoning length for Qwen3-32B is about 16,000 tokens. The third set, MMReD, is a synthetic benchmark for reasoning in dense contexts: it requires aggregating information across an entire sequence of steps (up to 128 steps) without using external tools. The leaderboard is part of a major update to the textual MERA and serves as a preview of the next platform release.