Roadmap

Build. Measure. Improve.

SZkrO is developed in revisions. Each revision has a defined purpose instead of growing as one giant uncontrolled experiment.

R0 — Model foundation

Separate project, Ollama runtime, candidate comparison, frozen V3 baseline and Granite 4.1 3B selected as the official base.

1

R1 — Training pipeline

Build SZkrO-specific dataset, validation and targeted LoRA/QLoRA training around instruction following, reasoning and coding.

2

R2+ — Iterative model development

Benchmark every revision, keep useful improvements, reject regressions and expand capabilities step by step.

N

SZkrO Native — long-term

Future separate line: own architecture/tokenizer/training pipeline and random-initialized weights, starting small before larger native models.