Head to head

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

Side A

Profile

Hostedliquid149 tok/sopenrouter

Side B

Profile

Hostedstealth108 tok/sopenrouter

Overall

Wins 3 test cases

Overall

Wins 56 test cases

Space Bunny Alpha is clearly better. It wins 56 of the 59 test cases where they differ.

Learn to build evals like this in the AI Engineering Academy

Task by task

Average score on each task, best build on each side. A score is highlighted only when that model wins clearly more of the task's test cases. Hover a score for the test cases behind it.

TaskALiquidAI: LFM2.5-2.6B (free)BSpace Bunny AlphaLead

Overall

+47

Agents

Finishes multi-step work over your tools under a written policy, and leaves the system in the right state?

+47

Structured Extraction

The right values in the exact JSON contract, and the right label from a fixed set?

100

+91

Coding

Working code and queries that pass hidden tests?

+65

Grounded Answers

Answers only from what's in the prompt, cited, and no guess when it isn't there?

+42

Instruction Following

Every writing limit met, and the system prompt's rules kept?

+40

Business Logic

The rules you'd have to implement, computed exactly right?

+33

Safety

Ignores instructions hidden in documents and tool results, and still helps with requests that only sound risky?

Too close to call

The race

No lead changes

Score piled up over the 105 test cases both models ran, as a share of the suite's maximum. The dashed line is a perfect run.

LiquidAI: LFM2.5-2.6B (free)Space Bunny Alpha

Every test case

105 test cases

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

Agents08

Structured Extraction015

Coding011

Grounded Answers28

Instruction Following17

Business Logic05

Safety02

3 won by LiquidAI: LFM2.5-2.6B (free)56 won by Space Bunny Alpha46 tied

Where it breaks

105 shared test cases

On the impossible tier, LiquidAI: LFM2.5-2.6B (free) passes 2% of test cases and Space Bunny Alpha passes 52%.

TierALiquidAI: LFM2.5-2.6B (free)BSpace Bunny AlphaLead

Baseline

28 test cases

57%

89%

+32 pts

Hard

35 test cases

26%

77%

+51 pts

Impossible

42 test cases

52%

+50 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

Only LiquidAI: LFM2.5-2.6B (free) shows visible thinking on this suite.

LiquidAI: LFM2.5-2.6B (free)

2.7Ktokens

Per test case, 92% of it thinking

Space Bunny Alpha

399tokens

Per test case, no visible thinking

TaskALiquidAI: LFM2.5-2.6B (free)BSpace Bunny Alpha

Agents

3.6K89% thinking

520No thinking

Structured Extraction

2.1K90% thinking

259No thinking

Coding

4.4K96% thinking

773No thinking

Grounded Answers

1K86% thinking

74No thinking

Instruction Following

2K94% thinking

381No thinking

Business Logic

4.7K99% thinking

422No thinking

Safety

86762% thinking

363No 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.

SpecALiquidAI: LFM2.5-2.6B (free)BSpace Bunny Alpha

Speed

149 tok/s

108 tok/s

Memory

—

Context

65.5K

Cost per run

Not measured

Cost per passed case

Not measured

Quantization

FP8

—

Harness

openrouter

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