Head to head

Any two models from the board, on the same tasks and graders.

Side A

Profile

Hosteddeepseek215 tok/sopenrouter

Side B

Profile

Hostedopenai45 tok/sopenai

Overall

Wins 9 tasks, 38 test cases

Overall

Wins 0 tasks, 5 test cases

DeepSeek V4.1 Flash leads by 12 points and wins 9 of 9 tasks.

Learn to build evals like this in the AI Engineering Academy

Task by task

DeepSeek V4.1 Flash wins 9 of 9

Average score on each task, best build on each side. The higher score is highlighted. Hover a score for the test cases behind it.

TaskADeepSeek V4.1 FlashBQwen3.8 27BLead

Overall

+12

Tool Calling

Right tool, right args, right sequence?

+13

Structured Output

Valid JSON matching the schema, correct fields?

RAG / Retrieval QA

A correct answer grounded in the given context, with nothing made up?

Context Recall

Finds the one fact buried in a long context?

100

Coding

Working code that passes hidden unit tests?

+34

Reasoning & Math

Multi-step math and logic with the exact right answer?

Instruction Following

Obeys format, length, and content constraints to the letter?

+21

Classification

Correct label from a fixed set?

+17

Summarization

Compact summary that keeps the facts and invents none?

The race

No lead changes

Score piled up over the 238 test cases both models ran, as a share of the suite's maximum. The dashed line is a perfect run.

DeepSeek V4.1 FlashQwen3.8 27B

Every test case

238 test cases

One square per test case, in suite order. A square takes the color of the model that scored higher on it.

Tool Calling72

Structured Output10

RAG / Retrieval QA43

Context Recall10

Coding90

Reasoning & Math20

Instruction Following60

Classification50

Summarization30

38 won by DeepSeek V4.1 Flash5 won by Qwen3.8 27B195 tied

Where it breaks

238 shared test cases

On the impossible tier, DeepSeek V4.1 Flash passes 44% of test cases and Qwen3.8 27B passes 29%.

TierADeepSeek V4.1 FlashBQwen3.8 27BLead

Baseline

36 test cases

94%

86%

+8 pts

Hard

69 test cases

67%

57%

+10 pts

Impossible

133 test cases

44%

29%

+15 pts

Tiers come from the suite. Baseline is fair, hard is adversarial, and impossible is built so that nothing solves it. Each number is the share of test cases the model fully passed.

Thinking spend

Both models spend about the same on thinking per test case.

DeepSeek V4.1 Flash

3.2Ktokens

Per test case, 97% of it thinking

Qwen3.8 27B

3.4Ktokens

Per test case, 98% of it thinking

TaskADeepSeek V4.1 FlashBQwen3.8 27B

Tool Calling

3.7K84% thinking

4.8K88% thinking

Structured Output

2.3K99% thinking

2.6K99% thinking

RAG / Retrieval QA

2.8K100% thinking

2.5K100% thinking

Context Recall

59099% thinking

50398% thinking

Coding

6.3K96% thinking

6.2K100% thinking

Reasoning & Math

5.4K100% thinking

Instruction Following

2.9K99% thinking

2.9K100% thinking

Classification

2.5K100% thinking

2.7K100% thinking

Summarization

1.7K99% thinking

1.9K99% thinking

Average output tokens per test case, and how much of it was thinking.

What you'd run

The build behind each score. The better value in each row is highlighted.

SpecADeepSeek V4.1 FlashBQwen3.8 27B

Speed

215 tok/s

45 tok/s

Memory

—

Context

—

Cost per run

$0.7017

Quantization

—

Harness

openrouter

openai

Faster, smaller, cheaper and more context count as better. Quantization and harness only describe the builds.