llama.cpp Slower Than Ollama? Local LLM Deployment Config Pitfalls

anshulsingh8326 · reddit · 2026-08-09

A developer testing on an RTX 4070 (12GB) found that running the same Q4K quantized model in llama.cpp yielded a lower generation speed (55 TPS) compared to Ollama's default configuration (61 TPS).

The poster shared detailed launch arguments (e.g., -ngl 99, -fa, -ctk q80) and hardware specs to ask for help. This highlights how default parameter optimizations and context processing mechanisms across different local tools significantly impact performance, and manual tuning can easily lead to bottlenecks.

Original post →

More from Infra

Infra channel →