MegaCapybara releases RTX5090 inference engine claiming 2x speed

On October 3, 2026, a developer announced MegaCapybara on Reddit's r/LocalLLaMA. It is an inference engine purpose-built for the RTX5090 and currently focused on Qwen3.8 27B, with the author saying more models will come later. The engine code is published on GitHub and the weights on Hugging Face.

The central claim is speed. According to the author, MegaCapybara roughly doubles the decode speed of Ninfer, which until that point was the SOTA engine for the RTX5090, and the gain is said to hold for both single-task and multi-task concurrency. On numbers, the material states 500+ t/s for a single stream, bursts of up to 650 t/s, and up to 2600 t/s at 12 concurrent requests. The original post describes 2600 t/s as what happens "if stars align," while the report summary attributes that figure to the 12-concurrent case, so the two accounts do not line up exactly. MegaCapybara also supports 800k context and ships with VRAM-RAM-disk cache management, Loop Guard, and a GUI.

The story currently stops at the announcement. All performance figures come from the author; the material gives no test conditions, no third-party reproduction, and no independent benchmark, so whether MegaCapybara actually reaches twice the speed of Ninfer cannot be confirmed from the available information.