Qwen3.8-27B vs Gemma4-26B: RTX 4090 Benchmarks
Introduction I tested five current local vision-language models on one NVIDIA RTX 4090 across Ollama and direct llama.cpp, centered on Qwen3.8-27B and Gemma4-26B. The work covered installation, model downloads, coding-harness configuration, practical text and vision context ceilings, multi-token prediction (MTP), short and filled-context throughput, CPU offload, matched runtime performance, and 360 W versus 450 W GPU power limits. Two models emerged as the most useful daily pair: Use Gemma4-26B for the fastest generation and longest practical full-GPU context. Use Qwen3.8-27B for difficult coding, agents, reasoning, and more reliable visual interpretation. Use Ollama for normal daily operation; use direct llama.cpp when exact GGUF, K/V precision, MTP, projector, or benchmarking control matters. The complete setup and all five-model results are retained below so the conclusions remain reproducible. ...