xiaomi's mimo-v2.5 owns the ultra-long-context tier on openrouter – 31.6% of all 1m–10m token requests, ~4x the next model

weekly requests, 1m–10m tokens (prompt + completion) – @OpenRouter:

• mimo-v2.5 – @XiaomiMiMo – 31.6%
• gemini 3 flash preview – @GoogleDeepMind – 8.6%
• gpt-5.6 luna pro – @OpenAI – 8.1%
• glm 5.2 – @Zai_org – 6.9%
• gpt-5.6 sol pro – openai – 6.6%
• gpt-4o-mini – openai – 6.0%
• kimi k3 – @Kimi_Moonshot – 4.8%
• mimo-v2.5-pro – xiaomi – 3.1%
• gpt-5.6 terra pro – openai – 3.0%
• others – 21.5%

observations:

1. one model owns the 1m+ tier – mimo-v2.5 pulls 31.6% of all 1m–10m token requests, ~4x #2 (gemini 3 flash, 8.6%). at massive context, users overwhelmingly pick one model

2. wide, not deep – openai charts 4 separate models (luna, sol, terra pro + 4o-mini) for ~23.7% combined, yet none crack 9%. no single openai model is the default this far out

3. fat tail – "others" is 21.5%, bigger than every model except mimo. past the leader, ultra-long-context demand is heavily fragmented