LightSeek reports 685 tokens per second on Qwen with four GB300s

LightSeek says TokenSpeed produced the per-user result on four NVIDIA GB300 GPUs using NVFP4, but it has not disclosed prompt length, concurrency or other benchmark conditions.

By · Published

Primary source: Qwen / Alibaba

Why it matters

TokenSpeed's result puts an open-source serving engine in the conversation around Qwen deployment, while the undisclosed test setup limits what operators can infer from 685 tokens per second.

AI-generated illustration of a large industrial fan beside rows of server racks, with bundled cables running across the floor and blue lights glowing at their ends.

LightSeek Foundation reported that its TokenSpeed inference engine generated 685 output tokens per second per user with Alibaba's Qwen3.8-Flash-Next, running on four NVIDIA GB300 GPUs with NVFP4, according to the Foundation's post. Qwen amplified the result on October 14th in a congratulatory post.…

Reader comments

Conversation for this story loads after sign-in.