Xiaomi’s MiMo-V2.6-Flash leads CyberGym benchmark

CyberGym evaluates AI agents on cybersecurity tasks, including vulnerability identification, security analysis and controlled security challenges. LLM Stats tracks 20 models on the benchmark, which is scored on a 0-1 scale and categorized under safety, agents and code.

Xiaomi’s MiMo-V2.6-Flash leads the ranking with a score of 0.951, followed by MiMo-V2.6-Pro at 0.940 and DeepSeek-V4.1-Flash at 0.881. Other top-ranked systems include Atria Dawn Preview, Gemini 3.8 Flash Cyber, GLM-5.3, DeepSeek-V4-Pro-0813, Gemini 3.5 Flash Cyber, Claude Mythos Preview and GPT-5.5.

The leaderboard lists 0 verified results and 20 self-reported results, with an average score of 0.787 in the FAQ. MiMo-V2.6-Flash is also identified as the top open-source model and the cheapest model within 10% of the leader, priced at $0.14 per million input tokens. The CyberGym leaderboard was last updated in October 2026.