ninfer vs llama.cpp: Qwen3.8 27B Runs 2.3x Faster

On October 3, 2026, a user on Reddit's LocalLLaMA forum posted a hands-on comparison of ninfer and llama.cpp, both running Qwen3.8 27B on an RTX 5090. The test used a single oneshot run with thinking enabled at xhigh, a 120k context window, default sampling settings, and the same prompt for both backends — described by the poster as covering physics and spin with the full rule set.

According to the post summary, ninfer decoded at roughly 147 tokens/s, while llama.cpp with Q4_K_M and MTP reached about 141 tokens/s. Wall-clock time came in at 7 minutes 40 seconds for ninfer versus 8 minutes 13 seconds for llama.cpp. The poster also said that switching to ninfer's NVFP4 build produced a larger output.

One point of disagreement is worth flagging: the headline says ninfer is 2.3x faster than llama.cpp, but the decode figures in the summary differ by only about 4%. The excerpt shown is also incomplete — it cuts off in the middle of a table row that lists ninfer's precision for the non-NVFP4 build alongside MTP. That leaves the underlying numbers only partly verifiable from the material available.

So far this remains a single user-generated benchmark post. No other reports, follow-up tests, or official statements from either project have appeared.