[mle]/DaBench

MiMo-V2.6-Flash vs Underdog Saluki 27B

Side A
Model page
Side B
Model page
Overall
45
vs
Overall
33

MiMo-V2.6-Flash is clearly better. It wins 35 of the 52 test cases where they differ.

Learn to build evals like this in the AI Engineering Academy

Task by task

A score turns orange only when that model wins clearly more of the task's test cases. Under each score, the average time per test case on the machine it ran on. Underdog Saluki 27B spends 1.5× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
451m 21s
4.2K, 92% thinking6.9K, 84% thinking
335m 15s
Agents
381m 19s
2.2K, 86% thinking11.8K, 97% thinking
409m 2s
Structured Extraction
6238.0s
2.6K, 91% thinking4.4K, 92% thinking
433m 17s
Coding
242m 40s
9.9K, 97% thinking9.2K, 49% thinking
166m 54s
Grounded Answers
6242.1s
2.6K, 89% thinking4.5K, 92% thinking
433m 22s
Instruction Following
641m 2s
4.8K, 90% thinking8K, 94% thinking
466m 5s
Business Logic
4356.4s
4.4K, 96% thinking5.2K, 81% thinking
333m 55s
Security & permissions
662m 6s
3.8K, 88% thinking4.9K, 91% thinking
433m 47s
Diagnosis & investigation
01m 33s
1.5K, 83% thinking8.7K, 94% thinking
06m 39s

What you'd run

The build behind each score. Orange marks the faster, smaller, cheaper or longer-context side.

SpecAB
Speed
66 tok/s
23 tok/s
Machine
Hosted
Apple M5 Pro, 48 GB
Memory at 32K
—
10.2 GB
Context
1.1M (better)
256K
Cost per run
$0.33
Not measured
Cost per case
$0.0021
Not measured
Time per case
1m 21s (better)
5m 15s
Quantization
BF16
IQ2_XS
Harness
openrouter
llama.cpp

The race

4 lead changes

Each model's running score as the test cases go by. The dashed line is a perfect run.

MiMo-V2.6-FlashUnderdog Saluki 27B
025507510050100150Perfect run4533

Every test case

153 test cases

Each square is one test case, colored by the model that scored higher.

35 won by MiMo-V2.6-Flash17 won by Underdog Saluki 27B101 tied