RTX 3090 Reaches 382 Tokens per Second With Qwen3.8-27B Optimizations

A developer reported a sharp increase in token generation speed for Qwen3.8-27B running on NVIDIA’s RTX 3090 after using a custom inference engine. The GPU produced 82 tokens per second in the initial setup and 382 tokens per second with the engine, a roughly 4.6-fold increase, according to Wccftech.

The RTX 3090 has 24GB of VRAM and memory bandwidth of 936.2GB/s. Those specifications can support prompt processing and token generation, but the card’s memory capacity is a constraint for running larger language models, the report says.

The speed gain came from inference optimizations rather than a hardware change. The source describes the result as close to five times faster, while noting that the process had some hiccups. It also cautions that token generation can become much faster when Qwen3.8-27B is repeating text it has already produced.

The figures are tied to this developer’s setup and should not be read as a general benchmark for every RTX 3090 system. The source does not provide further details about the tests or the custom engine in the supplied material.