Does Strata like projects trade accuracy for speed?
retained text
I am tired to answer on people questions like "How you gain speed? Your model output is shit, it has to have low quality." or something like that. So I will answer to that with benchmark. So I ran it against llama.cpp with the same Qwen3.8-Flash-Next IQ2\_XS file, same prompt, and the same 128-token reply. This was on my RTX 3060 12 GB / Ryzen 7600X machine. I did six warm runs for each engine. ||RustingStrata|llama.cpp| |:-|:-|:-| |Writing speed|**33.0 tok/s**|25.6 tok/s| |Range across runs|29.3–35.8|23.2–26.9| |Reading the first prompt|**109 tok/s**|31 tok/s| So RustingStrata was about 29% faster at writing in this test. Speed is nice, but I did not want to claim a win without checking what it does to the model’s output. I also ran both engines over the same 9,180 WikiText-2 tokens. |Quality check|Result| |:-|:-| |Perplexity: RustingStrata|2.9212| |Perplexity: llama.cpp|2.9156| |Same most-likely next token|96.78%| |Mean KLD|0.0098| The perplexity gap is tiny, small enough to be noise here. I do not see RustingStrata getting its speed by losing measurable quality, at least with this model and this GPU. This is not a claim that it wins on every model or every PC. It is one model, one quantization and one card. If anyone has any other questions about this, I am happy to answer it. Repo: [RustingStrata](https://github.com/RustingStudio/RustingStrata)
document fetched 2026-11-11 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xsw9mr/does_strata_like_projects_trade_accuracy_for_speed