[mle]/DaBench

GPT-6 Luna vs Underdog Saluki 27B

Side A
Model page
Side B
Model page
Overall
61
vs
Overall
33

GPT-6 Luna is clearly better. It wins 58 of the 67 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. Underdog Saluki 27B spends 4.3× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
6149.1s
1.9K, 73% thinking6.9K, 84% thinking
335m 15s
Agents
561m 18s
1.3K, 67% thinking11.8K, 97% thinking
409m 2s
Structured Extraction
9514.5s
999, 72% thinking4.4K, 92% thinking
433m 17s
Coding
732m 18s
3.6K, 60% thinking9.2K, 49% thinking
166m 54s
Grounded Answers
6715.6s
1.1K, 72% thinking4.5K, 92% thinking
433m 22s
Instruction Following
8432.3s
2K, 74% thinking8K, 94% thinking
466m 5s
Business Logic
3743.1s
3.2K, 93% thinking5.2K, 81% thinking
333m 55s
Security & permissions
6224.5s
1K, 68% thinking4.9K, 91% thinking
433m 47s
Diagnosis & investigation
1737.7s
1.1K, 68% thinking8.7K, 94% thinking
06m 39s

What you'd run

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

SpecAB
Speed
90 tok/s
23 tok/s
Machine
Hosted
Apple M5 Pro, 48 GB
Memory at 32K
—
10.2 GB
Context
1.1M (better)
256K
Cost per run
$0.10
Not measured
Cost per case
$0.00067
Not measured
Time per case
49.1s (better)
5m 15s
Quantization
—
IQ2_XS
Harness
openrouter
llama.cpp

The race

1 lead change

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

GPT-6 LunaUnderdog Saluki 27B
025507510050100150Perfect run6133

Every test case

153 test cases

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

58 won by GPT-6 Luna9 won by Underdog Saluki 27B86 tied