[mle]/DaBench

Beam 1.0 501B A23B vs DeepSeek V4.1 Flash

Side A
Model page
Side B
Model page
Overall
57
vs
Overall
64

DeepSeek V4.1 Flash is clearly better. It wins 32 of the 43 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 spends 1.5× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
5717.2s
4K, 87% thinking6K, 91% thinking
643m 35s
Agents
5521.8s
4.8K, 90% thinking6.1K, 82% thinking
464m 42s
Structured Extraction
819.98s
3.4K, 91% thinking4.4K, 94% thinking
862m 8s
Coding
4819.9s
4.5K, 70% thinking9.7K, 95% thinking
514m 55s
Grounded Answers
6911.6s
2.7K, 88% thinking3.9K, 90% thinking
762m 17s
Instruction Following
7820.9s
5.6K, 90% thinking7.5K, 94% thinking
894m 24s
Business Logic
5714.0s
4.2K, 95% thinking7K, 97% thinking
814m 1s
Security & permissions
5120.6s
3.2K, 88% thinking4.4K, 81% thinking
693m 3s
Diagnosis & investigation
1724.3s
3.1K, 90% thinking3.2K, 67% thinking
172m 8s

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)
30 tok/s
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M
1M
Cost per run
$0.83
$0.39 (better)
Cost per case
$0.0054
$0.0026 (better)
Time per case
17.2s (better)
3m 35s
Quantization
FP8
FP8
Harness
openrouter
openrouter

The race

2 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
025507510050100150Perfect run5764

Every test case

153 test cases

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

11 won by Beam 1.0 501B A23B32 won by DeepSeek V4.1 Flash110 tied