Model Drop

MiMo-V2.6-Pro

September 21, 2026 2 min read

MiMo-V2.6 Pro, which Xiaomi released and open-sourced on September 21, 2026 (September 22 in Beijing, the date on its launch post), is the flagship of its MiMo-V2.6 series: a 1.02T-parameter omnimodal mixture-of-experts (42B active) that takes text, images, video and audio.

Model: MiMo-V2.6 Pro (XiaomiMiMo/MiMo-V2.6-Pro-RL)
Maker: Xiaomi
Ship date: September 21, 2026
Pricing: $0.435 / $0.87 per M tokens (MiMo API); free weights
Context: 1M tokens
Weights: Open (MIT)
Where you can use it:
- Hugging Face and ModelScope weights
- MiMo API, AI Studio, MiMo Code, MiMo Desktop
When to reach for it:
- Long agent runs mixing text, image, audio
- Strong agent work at a low token price
- Skip it for exploit-writing security work
How it compares:
- Open rival to Kimi K3 and GLM-5.3
- Close to Claude Opus 5 on agent tasks
What makes it different:
- RL run livestreamed, code and envs open
- UltraSpeed mode, up to 20x faster output
Early reactions: Hugging Face's Thomas Wolf praised an "impressive level of openness on such a large run," and coverage singled out the public RL bill and MIT weights.
Read more: Introducing MiMo-V2.6 series, model card, technical report

The benchmarks

Xiaomi's tech report runs every rival with a configurable reasoning effort at its highest setting (max) but doesn't state MiMo's own setting; the scores come from its September 22 benchmark sheet, comparing against open Kimi K3 and closed Claude Opus 5, GPT-5.6 Sol and Fable 5, with a dash where a rival has no result. Kimi K3's scores appear only in the blog's appendix table, not the model card, and GDPval-AA is the Elo that Artificial Analysis reports.

Xiaomi's numbers

DeepSWE v1.1 (long software tasks): 71.9%. +2.9 over Kimi K3 (69.0%), −2.1 under Claude Opus 5 (74.0%), +1.9 over Fable 5 (70.0%).
Terminal-Bench 2.1 (agentic terminal tasks): 89.9%. +1.6 over Kimi K3 (88.3%), +0.8 over Claude Opus 5 (89.1%), +1.1 over GPT-5.6 Sol (88.8%).
Terminal-Bench 4.0 (harder terminal tasks): 34.9%. +22.3 over Kimi K3 (12.6%), −14.1 under Claude Opus 5 (49.0%), −5.0 under GPT-5.6 Sol (39.9%).
OSWorld-Verified (computer use): 82.0%. −2.8 under Kimi K3 (84.8%), −1.4 under Claude Opus 5 (83.4%), −1.0 under GPT-5.6 Sol (83.0%).
Toolathlon-Verified (multi-app tool use): 76.9%. +0.4 over Kimi K3 (76.5%), −3.7 under Claude Opus 5 (80.6%), +2.0 over GPT-5.6 Sol (74.9%).
AutomationBench (workflow automation): 53.1%. +6.4 over Kimi K3 (46.7%), +2.8 over Claude Opus 5 (50.3%), +7.3 over GPT-5.6 Sol (45.8%).
JobBench (professional job tasks): 62.0%. +7.7 over Kimi K3 (54.3%), −3.7 under Claude Opus 5 (65.7%), +16.6 over GPT-5.6 Sol (45.4%).
GDPval-AA (office-work Elo, v2.1): 1673. +149 over Kimi K3 (1524), −35 under Claude Opus 5 (1708), +85 over GPT-5.6 Sol (1588).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 46.3. +2.7 over Kimi K3 (43.6), +1.5 over GLM-5.3 (44.8), +6.4 over Qwen3.8-2.4T-A95B (39.9), +6.8 over DeepSeek V4.1 Flash (39.5).