[mle]/DaBench

Beam 1.0 501B A23B vs Qwen3.8 27B

Side A
Model page
Side B
Model page
Overall
57
vs
Overall
48

Beam 1.0 501B A23B is clearly better. It wins 45 of the 66 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. Qwen3.8 27B spends 1.5× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
5717.2s
4K, 87% thinking5.9K, 91% thinking
4833.3s
Agents
5521.8s
4.8K, 90% thinking5.2K, 78% thinking
5736.6s
Structured Extraction
819.98s
3.4K, 91% thinking4.6K, 92% thinking
7320.6s
Coding
4819.9s
4.5K, 70% thinking9.3K, 94% thinking
3753.0s
Grounded Answers
6911.6s
2.7K, 88% thinking4.2K, 91% thinking
6131.6s
Instruction Following
7820.9s
5.6K, 90% thinking6.9K, 93% thinking
6733.6s
Business Logic
5714.0s
4.2K, 95% thinking7.6K, 98% thinking
3632.9s
Security & permissions
5120.6s
3.2K, 88% thinking3.7K, 85% thinking
5122.6s
Diagnosis & investigation
1724.3s
3.1K, 90% thinking5.9K, 85% thinking
040.0s

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)
196 tok/s
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M (better)
256K
Cost per run
$0.83 (better)
$1.28
Cost per case
$0.0054 (better)
$0.0083
Time per case
17.2s (better)
33.3s
Quantization
FP8
FP8
Harness
openrouter
openrouter

The race

1 lead change

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

Beam 1.0 501B A23BQwen3.8 27B
025507510050100150Perfect run5748

Every test case

153 test cases

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

45 won by Beam 1.0 501B A23B21 won by Qwen3.8 27B87 tied