R0 · frozen V3

Benchmark before training.

24 cases across Polish, instruction following, reasoning, coding, code analysis and context. This is the frozen pre-training reference point for future SZkrO revisions.

87.5%
Correctness
66.7%
Instruction adherence
62.5%
Strict pass
15.5
Median tok/s

Category correctness

Polish
4/4
Code analysis
4/4
Context
4/4
Instruction
3/4
Reasoning
3/4
Coding
3/4
R0 comparison

Why Granite won.

ModelCorrectnessInstructionStrictMedian caseMedian tok/s
Granite 4.1 3B87.5%66.7%62.5%4.1 s15.5
Gemma 3 4B83.3%62.5%54.2%5.9 s10.2
Gemma 4 E2B79.2%79.2%62.5%40.1 s9.5
Phi-4 Mini 3.8B66.7%62.5%45.8%5.4 s12.5
Llama 3.2 3B66.7%62.5%50.0%4.6 s11.9