0:00 / 0:00

With a 3090, I drove Qianwen 3.8-27B on vLLM to 64 concurrent channels, and the total exceeded 1,000 tok/s.

Name: With a 3090, I drove Qianwen 3.8-27B on vLLM to 64 concurrent channels, and the total exceeded 1,000 tok/s.
Uploaded: 2026-10-28T01:32:34.000Z
Duration: 55 s
Description: Open source project HyperQwen: single card 24G, 3090 limited to 250W actual test. Single-person chat takes about 127 tok/s, 64 concurrent conversations total about 1035 tok/s, and the maximum number of contexts can be opened is 150,000.

Open source project HyperQwen: single card 24G, 3090 limited to 250W actual test. Single-person chat takes about 127 tok/s, 64 concurrent conversations total about 1035 tok/s, and the maximum number of contexts can be opened is 150,000. It can be started with two commands: docker compose --profile single or batch.

Qwen AI FansOct 281.5K

Keep scrolling for the next clip ↓All of qwen38 Archive