[mle]/DaBench

Qwen3.8 Flash Next vs Spark X2.6 4B

Side A
Model page
Side B
Model page
Overall
42
vs
Overall
15

Qwen3.8 Flash Next is clearly better. It wins 62 of the 71 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. Spark X2.6 4B spends 2.1× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
421m 15s
4.1K, 80% thinking8.2K, 84% thinking
151m 3s
Agents
522m 4s
4.6K, 78% thinking11.5K, 97% thinking
141m 32s
Structured Extraction
5042.2s
3.1K, 90% thinking5.3K, 78% thinking
1039.5s
Coding
521m 36s
6.2K, 61% thinking9.1K, 48% thinking
41m 8s
Grounded Answers
5147.0s
3.1K, 88% thinking6.6K, 86% thinking
1050.2s
Instruction Following
621m 30s
5.2K, 88% thinking11.9K, 90% thinking
371m 29s
Business Logic
251m 0s
3.9K, 90% thinking4.9K, 86% thinking
1036.3s
Security & permissions
451m 8s
2.8K, 82% thinking7K, 94% thinking
3154.3s
Diagnosis & investigation
01m 15s
3.4K, 69% thinking11.6K, 98% thinking
01m 30s

What you'd run

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

SpecAB
Speed
68 tok/s
136 tok/s
Machine
Hosted
RTX 5090, 32 GB
Memory at 32K
—
9.4 GB
Context
1M (better)
605K
Cost per run
$0.46 (better)
$1.61 est.
Cost per case
$0.003 (better)
$0.011 est.
Time per case
1m 15s
1m 3s (better)
Quantization
—
BF16
Harness
openrouter
llama.cpp

The race

No lead changes

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

Qwen3.8 Flash NextSpark X2.6 4B
0255075100501001504215

Every test case

153 test cases

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

62 won by Qwen3.8 Flash Next9 won by Spark X2.6 4B82 tied