Head to head

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

Side A

Profile

Hostedpoolside48 tok/sopenrouter

Side B

Profile

Hostednvidia21 tok/sopenrouter

Overall

Wins 1 task, 23 test cases

Overall

Wins 8 tasks, 60 test cases

Nemotron 3 Ultra leads by 15 points and wins 8 of 9 tasks.

Learn to build evals like this in the AI Engineering Academy

Task by task

Nemotron 3 Ultra wins 8 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.

TaskALaguna S 2.1BNemotron 3 UltraLead

Overall

+15

Tool Calling

Right tool, right args, right sequence?

+14

Structured Output

Valid JSON matching the schema, correct fields?

+11

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?

+28

Coding

Working code that passes hidden unit tests?

+20

Reasoning & Math

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

+21

Instruction Following

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

+25

Classification

Correct label from a fixed set?

+24

Summarization

Compact summary that keeps the facts and invents none?

+15

The race

2 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.

Laguna S 2.1Nemotron 3 Ultra

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 Calling116

Structured Output14

RAG / Retrieval QA34

Context Recall05

Coding514

Reasoning & Math06

Instruction Following07

Classification18

Summarization26

23 won by Laguna S 2.160 won by Nemotron 3 Ultra155 tied

Where it breaks

238 shared test cases

On the impossible tier, Laguna S 2.1 passes 19% of test cases and Nemotron 3 Ultra passes 26%.

TierALaguna S 2.1BNemotron 3 UltraLead

Baseline

36 test cases

53%

78%

+25 pts

Hard

69 test cases

29%

55%

+26 pts

Impossible

133 test cases

19%

26%

+7 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

Laguna S 2.1 spends 1.3× as many thinking tokens per test case.

Laguna S 2.1

2.5Ktokens

Per test case, 76% of it thinking

Nemotron 3 Ultra

2.6Ktokens

Per test case, 56% of it thinking

TaskALaguna S 2.1BNemotron 3 Ultra

Tool Calling

4.2K74% thinking

4.3K47% thinking

Structured Output

1.4K97% thinking

1.5K47% thinking

RAG / Retrieval QA

1.5K79% thinking

1.9K62% thinking

Context Recall

66683% thinking

1K65% thinking

Coding

2.9K42% thinking

5K69% thinking

Reasoning & Math

5.6K87% thinking

3.6K54% thinking

Instruction Following

1.8K34% thinking

2.2K47% thinking

Classification

2.8K100% thinking

1.6K57% thinking

Summarization

51393% thinking

1.4K59% 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.

SpecALaguna S 2.1BNemotron 3 Ultra

Speed

48 tok/s

21 tok/s

Memory

—

Context

262.1K

Cost per run

Quantization

—

Harness

openrouter

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