# Llama.cpp performance improvement over time (Gemma4 models) — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/a8e557ed8a91ed66

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-15 20:34 ingested 2026-11-16 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xx607x/llamacpp_performance_improvement_over_time_gemma4/)

## retained text



```
Few months ago - after Ninfer came out - I started a little hobby project to vibe-code my own engine for my hardware (RTX 5080 16GB Laptop) and check if it could beat llama.cpp - since at the time llama.cpp was kinda slow in comparison to something like vLLM, but vLLM is using VRAM very wastefully.

Back then it was quite easy to actually get something up and running faster than llama.cpp.

Today I just benchmarked llama.cpp again, and it caught up / surpassed my vibe-coded engine. I prefer using a stable engine with proper features than all the vibe-coded ones, including my own. So it is really nice to see that the project - while a bit slower - lands significant performance improvements over time.

I put together a little chart of benchmarks I had laying around which visualizes this quite nicely. It is by no means scientific - it's always a fixed 16k prefill 1k output over Wikipedia text with the optimal settings for each engine. The 12B model is the same checkpoint, for the 26B model these are different quants for all engines, so not even apples-to-apples and llama.cpp wins.
```



document fetched 2026-11-16 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1xx607x/llamacpp_performance_improvement_over_time_gemma4
