Experimental studies of local AI inference: throughput, memory residency, context scaling and application latency.
Qwen3.8-Flash-Next
Hardware: EVGA GeForce RTX 3090 OC FTW3 Ultra, Ryzen 9 8945HX, nominal 96 GB RAM.
- Interactive report on GitHub Pages
- Comprehensive Markdown report
- Measurements CSV
- Configurations and data JSON
The English report starts with three tested Q3/upstream llama.cpp configurations: 32k interactive text, 256k text and the retained 256k integrated serving profile. It provides exact startup commands, pinned versions, build flags, CPU/GPU placement, KV/batch/MTP settings and request-level parameters beside their measured results.
The tuning checklist explains what improved performance on this platform and which settings should not be copied between workloads. Q4/Strata is a separate experimental comparison, not the serving recommendation.
The HTML presents the research findings, interactive figures and a separate measurement explorer. Download index.html for offline use; no external runtime libraries are required.
Supplementary data include individual measurements, version-specific commands, recorded environments and artifact hashes. Complete raw input fixtures are not distributed; experimental scope and reproducibility constraints are described in the paper.