MiMo
by Xiaomi · Model · Language & code models
strong
Top open model on several independent leaderboards, though Kimi K3.1 now leads agents
#1 chat (open weights)
MiMo-V2.6-Distill-Qwen-9B is Xiaomi MiMo’s 9B model for code and agent workflows
AccessOpen weightsNeeds~573 GB; smallest build ~6.6 GBDownload573 GB for Pro weights; 6.6 GB via Ollama
Why we say so
Arena ranked MiMo 2.6 first among open weights on text, covering coding, writing and chat, ahead of GLM-5.3 and Kimi K3. It slipped from first to third on Arena agent, behind Kimi K3.1 and Inkling-2-Small. Artificial Analysis also puts it ahead of Mistral Large 4 and Claude Haiku 5.5 on intelligence per cost, and Arena's Code Arena has the Flash version beating Opus 4.7. No hands-on reports are on record yet.
Ranks #3 for agents (open weights)
Evidence
- Nov 15#3 on Arena agent among open weights
- Nov 14MiMo-V2.6-Flash outscored Opus 4.7 in Code Arena; across 105 models, beating April's top model cost 20x less by October
- Nov 11#2 on Arena agent among open weights
- Nov 11#1 on Arena text among open weights (1534)
- Nov 11#1 on Arena text among open weights (1469)
Show all 11
- Nov 11#1 on Arena text among open weights (1492)
- Oct 28Took #1 on Arena agent among open weights
- Sep 27MiMo 2.6 Pro (~1T params, 573 GB) and Flash (311B, 178 GB) released as open weights
- Sep 27Top open model on Artificial Analysis index, one point above GLM 5.3, at 13 cents per task
- Sep 27Pro close to GPT-6 Astra and above Fable 5 on DeepSWE; ties Grok 4.7 on the AA index
- Sep 20Pro model DeepSWE score rose from 58 to 67 during RL, close to top open models