[mle]/DaBench

Mellum2.1 Thinking vs Mistral Large 4

Side A
Model page
Side B
Model page
Overall
9
vs
Overall
28

Mistral Large 4 is clearly better. It wins 42 of the 47 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. Mellum2.1 Thinking spends 1.5× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
91m 56s
10.3K, 63% thinking6.2K, 72% thinking
281m 56s
Agents
92m 3s
10.8K, 93% thinking795, no thinking
1126.1s
Structured Extraction
51m 36s
8.3K, 51% thinking288, no thinking
153.78s
Coding
52m 11s
11.3K, 38% thinking9.3K, 37% thinking
332m 48s
Grounded Answers
131m 24s
7.3K, 62% thinking8.1K, 89% thinking
382m 45s
Instruction Following
283m 10s
17.3K, 63% thinking12.9K, 84% thinking
693m 47s
Business Logic
01m 39s
9K, 50% thinking8.8K, 77% thinking
242m 40s
Security & permissions
81m 40s
9.1K, 75% thinking3.7K, 83% thinking
331m 9s
Diagnosis & investigation
01m 15s
6.8K, 98% thinking4.4K, 85% thinking
01m 19s

What you'd run

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

SpecAB
Speed
90 tok/s
56 tok/s
Machine
Apple M5 Pro, 48 GB
Hosted
Memory at 32K
12.9 GB
—
Context
128K
512K (better)
Cost per run
$2.46 est. (better)
$2.67
Cost per case
$0.016 est. (better)
$0.017
Time per case
1m 56s
1m 56s
Quantization
Q8_0
—
Harness
llama.cpp
openrouter

The race

1 lead change

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

Mellum2.1 ThinkingMistral Large 4
025507510050100150Perfect run928

Every test case

153 test cases

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

5 won by Mellum2.1 Thinking42 won by Mistral Large 4106 tied