CL-SN55 941 —
GPT-61S 885 -0.6%
GPT-56T 861 —
GPT-6S 825 -1.4%
CL-OP55X 819 -1.4%
MUSE-SPK 793 -1.4%
QWEN-38X 779 -2%
GPT-6A 778 -0.5%
CL-OP5H 712 -1.7%
CL-OP5X 710 -1.7%
KIMI-K3X 700 -2.5%
CL-OP46H 695 -1.7%
CL-FAB5H 693 -1.7%
CL-OP55H 693 -1.7%
CL-OP47H 684 -1.7%
GEM-37FH 649 -1.8%
GEM-38FH 634 -3.4%
GPT-56S 574 -0.5%
CL-OP47 540 -0.6%
INKL 531 —
GEM-3P 494 —
CL-OP46 490 —
CL-OP48 483 —
GEM-31P 459 -5.4%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 -0.6%
GPT-56T 861 —
GPT-6S 825 -1.4%
CL-OP55X 819 -1.4%
MUSE-SPK 793 -1.4%
QWEN-38X 779 -2%
GPT-6A 778 -0.5%
CL-OP5H 712 -1.7%
CL-OP5X 710 -1.7%
KIMI-K3X 700 -2.5%
CL-OP46H 695 -1.7%
CL-FAB5H 693 -1.7%
CL-OP55H 693 -1.7%
CL-OP47H 684 -1.7%
GEM-37FH 649 -1.8%
GEM-38FH 634 -3.4%
GPT-56S 574 -0.5%
CL-OP47 540 -0.6%
INKL 531 —
GEM-3P 494 —
CL-OP46 490 —
CL-OP48 483 —
GEM-31P 459 -5.4%
GPT-6L 436 —
← Back to feed

Xiaomi MiMo-V2.6-Pro Breaks Top 10 in Code Arena — Highest MIT-License Open-Source Rank

Xiaomi’s MiMo-V2.6-Pro entered Arena.ai’s Code Arena WebDev leaderboard on September 28 at a top-10 position overall, placing third among MIT-licensed open-source models. The model was released on September 22 with fully open weights and a technical report — one of the few frontier-tier models released under a permissive license with a simultaneous open-source publication.

What Changed From V2.5

MiMo-V2.6 is built around what Xiaomi calls a recursive self-improvement (RSI) path: the training process scales reinforcement learning compute on verifiable complex tasks, allowing the model to expand capability through ongoing exploration and feedback rather than fixed curriculum.

The V2.6 series includes:

  • MiMo-V2.6-Pro — full flagship model, open weights + technical report, MIT license
  • MiMo-V2.6-Flash — smaller, faster variant, also open-sourced under MIT
  • MiMo-V2.6-Distill-Qwen-9B — distilled from V2.6 using RL research resources, also released

API prices for V2.6 are unchanged from V2.5, where Xiaomi cut prices by up to 86% in May. A desktop client and membership subscription plan launched simultaneously with the model, providing access to both Pro and Flash via Xiaomi’s consumer platform.

Arena Debut: Code WebDev and Vision

Arena.ai added MiMo-V2.6-Flash and MiMo-V2.6-Pro to Vision Arena on September 28. On the same day, V2.6-Pro’s Code Arena WebDev ranking was confirmed in the top 10 overall — in the prior V2.5 cycle, Xiaomi’s models competed primarily on efficiency; V2.6-Pro now competes on absolute capability, placing in the top 3 among MIT-licensed models specifically.

Arena’s Code Arena WebDev leaderboard ranks models on live web development tasks with real-world human votes. It is among the harder Arena verticals for open-source models to place highly, where most top positions have historically been held by closed proprietary systems.

For context on where Xiaomi stands: V2.5-Pro reached Intelligence Index 54 on Artificial Analysis in April, and the V2.6-Pro was reported to tie Grok 4.7 at AA Intelligence Index 46 at launch — a slight regression on the blended AA metric, but the Arena Code result suggests strong coding-specific capability improvement.

MIT License Significance

The MIT license on MiMo-V2.6-Pro and Flash makes them commercially deployable without restriction. Combined with the open technical report and RL research resources from the Distill-Qwen-9B release, Xiaomi is making a deliberate push to establish a research footprint outside consumer hardware. The move mirrors Alibaba’s Qwen strategy but from a device-first company with a strong retail distribution network.

The next threshold: whether MiMo-V2.6-Pro earns an SWE-bench Verified run. The V2.5-Pro was scored at 78.9% in May; a V2.6 run would test whether the RSI training path produced measurable agentic coding gains.