[mle]/DaBench

Beam 1.0 501B A23B vs DeepSeek V4.1 Flash 1017

Side A
Model page
Side B
Model page
Overall
57
vs
Overall
56

Too close to call. DeepSeek V4.1 Flash 1017 wins 25 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. DeepSeek V4.1 Flash 1017 spends 1.4× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
5717.2s
4K, 87% thinking5.3K, 91% thinking
5626.0s
Agents
5521.8s
4.8K, 90% thinking3.3K, 71% thinking
5826.7s
Structured Extraction
819.98s
3.4K, 91% thinking4.2K, 93% thinking
8117.2s
Coding
4819.9s
4.5K, 70% thinking9.4K, 95% thinking
4840.8s
Grounded Answers
6911.6s
2.7K, 88% thinking4.4K, 91% thinking
7119.8s
Instruction Following
7820.9s
5.6K, 90% thinking5.5K, 91% thinking
7625.6s
Business Logic
5714.0s
4.2K, 95% thinking6.6K, 97% thinking
6726.6s
Security & permissions
5120.6s
3.2K, 88% thinking4.7K, 87% thinking
4826.6s
Diagnosis & investigation
1724.3s
3.1K, 90% thinking3.2K, 74% thinking
020.2s

What you'd run

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

SpecAB
Speed
274 tok/s (better)
243 tok/s
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M
1M
Cost per run
$0.83
$0.75 (better)
Cost per case
$0.0054
$0.0049 (better)
Time per case
17.2s (better)
26.0s
Quantization
FP8
—
Harness
openrouter
openrouter

The race

12 lead changes

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

Beam 1.0 501B A23BDeepSeek V4.1 Flash 1017
025507510050100150Perfect run5756

Every test case

153 test cases

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

22 won by Beam 1.0 501B A23B25 won by DeepSeek V4.1 Flash 1017106 tied