Welcome
My name is Siarhei Berdachuk.
I'm a seasoned IT professional with over 30 years of experience in software engineering, team leadership, and cutting-edge solution design. I specialize in Java development, AI integration with Spring AI, and high-performance backend systems.
This blog is where I share insights on clean code, test-driven development, scalable architectures, and practical applications of AI in enterprise environments.
Latest Articles
07-Nov-2026
Gemma-4 12B QAT on an RTX 4090: Qualbench at 128K
Gemma-4 12B vs Strata on 4090: qual 42/50 vs 50/50; decode ~51 vs ~90-180 t/s; 128K usable but Strata wins prod.
07-Nov-2026
Gemma-4 31B QAT on an RTX 4090: Qualbench, Speed, and a 3K Ceiling
Temporary vLLM Gemma-4 31B QAT on 1×4090: qualbench 37/40 (~37/50), ~41 t/s, hard 3072, not efficient for real work (need ≥32K, prefer 128K).
07-Nov-2026
Strata IQ3_S vs FreeToken on an RTX 4090: Qualbench Score and Speed
FreeToken FP8 47/50 vs Strata IQ3_S 50/50 (0.2.11, parallel=2 @ 596 s); Maya 41/50 @ 32K; Gemma-4 QAT ~37/50 @ 3072.
30-Oct-2026
Maya GLM-5.3-Flash on an RTX 4090: Qualbench, Speed, and a 32K Ceiling
Project Maya Maya-S24 on 1×4090: qualbench 41/50 in ~10 min, decode ~30 t/s, hard 32K context, vs Strata IQ3_S 50/50 and ~4-6× faster decode.