AD
Track all your projects in one dashboard. Get 📊stats, 🔥heatmaps and 👀recordings in one self-hosted dashboard.
uxwizz.com
#126

MiMo-V2.6-Flash

Xiaomi Release: 2026-09-22 Tested on: 2026-09-22 01:13 xiaomi/mimo-v2.6-flash::medium
309B total (15B active)MoEOpen source
(medium) (low) (none)

Summary

MiMo-V2.6-Flash scores 7.4 on AI BENCHY and ranks #126. It has 10.0 reliability, a 65.2% pass rate, $0.190 total cost, and 60.62s average response time.

What makes MiMo-V2.6-Flash unique: It stands out most in Domain specific, where it ranks #4, while Anti-AI Tricks is its weakest area at #16. It uses unusually many reasoning tokens, which can help explain its slower or more expensive runs.

Model facts

Researched on 2026-09-22

Reported
Parameters
309B total (15B active)
Architecture
MoE
Availability
Open source
License
MIT

Official RL checkpoint card reports the parameter counts, MIT licensing, public weights, and SGLang/vLLM deployment instructions. The card identifies availability through OpenRouter.

Score

7.4

Consistency

7.8

Total Output Tokens

615,978

Total Input Tokens

121,240

Input Price

$0.140 / 1M

Output Price

$0.280 / 1M

Tests Correct

Wrong Tests: 11

Attempt pass rate: 65.2%

Flaky tests

6

Flaky tests had mixed outcomes across runs (at least one pass and one fail).

Response Time (avg)

60.62s

Response Time (max): 357.28s

Response Time (total): 1333.53s

Hamster playing table tennis

Prompt: Create a detailed SVG illustration of a hamster playing table tennis.

#126 MiMo-V2.6-Flash

medium
Cost
$0.004
Time
61.0s
Tokens
10,798 tok

Charts

Choose the first model, then click a second model to open a side-by-side page.

Total Output Tokens

Score vs Total Output Tokens

Quick Compare

Category Breakdown

Category Score Consistency Tests Correct
Anti-AI Tricks 5.9 4.1
Coding 7.2 5.1
Combined 7.1 9.1
Data parsing and extraction 10.0 10.0
Domain specific 5.3 10.0
General Intelligence 4.1 2.7
Instructions following 9.9 10.0
Puzzle Solving 7.8 10.0
Tool Calling 10.0 10.0
Trivia 3.0 10.0

Compared models