R0 foundation complete · PRE-TRAINING

SZkrO
local AI,
built seriously.

Samodzielny Zespół z rozwojem Okresowym. Lokalny projekt modelu AI rozwijany etapami — z mierzalnymi benchmarkami, przejrzystym statusem i naciskiem na wydajność.

Pronunciation: “szkro” · Current base: Granite 4.1 3B
SZkrO brand identity
Official brand direction · 2026
Frozen pre-training baseline

Measured before training.

R0 established a fixed V3 benchmark so future SZkrO revisions can be compared against a real baseline instead of vibes.

87.5%
Correctness
Granite 4.1 3B baseline
62.5%
Strict pass
Correct + instruction compliant
4.1 s
Median case time
24-case V3 benchmark
15.5
Median tok/s
Local inference
In development

SZkrO 1.0 · 3B direction

SZkrO is currently built on Granite 4.1 3B as the selected R0 base. R1 will focus on targeted improvement of instruction following, reasoning and coding while protecting existing strengths in Polish, context handling and code analysis.

Current stagePRE-TRAINING
R0 statusCOMPLETED
Base modelGranite 4.1 3B
RuntimeOllama
Next revisionR1 training pipeline
What matters

Small model. Clear targets.

Local-first

Designed around practical local inference instead of assuming unlimited cloud compute.

Benchmark-driven

Every major model revision should beat or justify itself against a frozen evaluation baseline.

Transparent lineage

Current SZkrO development clearly states its Granite 4.1 3B base instead of hiding model ancestry.

R0 is done. R1 comes next.

The foundation is frozen. The next model step is a project-specific dataset and training pipeline focused on the weaknesses identified in R0.

View roadmap →