Breaking

Xiaomi's MiMo-V2.6 Open-Source Models Rank #5 in Agent Arena After Scaled Reinforcement Learning

October 1, 2026 at 14:58 EDT

Foundation Models
AI Agents
Open Source

Xiaomi has released MiMo-V2.6-Pro and MiMo-V2.6-Flash, open-source multimodal foundation models trained via large-scale reinforcement learning that now rank among the strongest agentic systems, with MiMo-V2.6-Pro claiming fifth position on Agent Arena's open-source leaderboard. The Pro variant achieved a 3.17% net improvement across 8,158 real-world agentic sessions, representing a substantial jump from its predecessor's ranking position.

Agentic AI · October 2026

MiMo-V2.6: Open-Source Agentic Leap

Xiaomi's reinforcement-learning-trained models rank fifth in Agent Arena, outperforming on real-world tool use and user-correction handling.

Agent Arena Rank

Overall (Open-Source)

Steerability

All Models (All Licenses)

Net Improvement

+3.17%

8,158 Sessions

Training Cost

$3.5M

RL (6 Days)

Pro vs. Flash Scale

1.02T

Pro Total

42B active

309B

Flash Total

15B active

Flash trades some performance for 3× lower compute requirements and API cost.

Training Method

→ 750K trajectories across coding, agentic tasks, vision, security

→ 30 RL optimization steps in single integrated run

→ 6 days wall-clock time (livestreamed publicly)

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.