Radar ·
Author reports batch serving of Qwen3.8-27B on a single RTX 3090
Self-reported account describes serving Qwen3.8-27B on an RTX 3090 capped at 250W, with about 1035 tok/s total across 64 simultaneous chats and 127 tok/s for a single chat.
According to the text, the author served Qwen3.8-27B on an RTX 3090 capped at 250W, handling 64 simultaneous chats at about 1035 tok/s total and 127 tok/s for one chat. The context would reach 150K, and the setup would take two Docker Compose commands. The figures are batch throughput for a 27B model on consumer GPU hardware.
The data is self-reported, and the excerpt does not give enough methodology, software version or measurement conditions to reproduce the result. To verify it, consult the original report indicated by Radar, compare it with other measurements and, if you want to apply the information, run your own test with synthetic data.
Original source: Autor relata serving do Qwen3.8-27B em lote em uma RTX 3090