Qwen3.8 hits 255 tok/s on a single RTX 5090
What happened
A Hacker News post on October 14 describes single-card inference performance for Qwen3.8. A team combined post-training, quantization and kernel specialization to run Qwen3.8-Flash-Next-NVFP4 at 255 tokens per second on one RTX 5090. The model targets coding and everyday tasks. The post also lists a single RTX6000 as the recommended hardware, where it supports 4 users at an aggregate 334 tokens per second. The multi-user figure and the RTX 5090 speed come from different hardware setups.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageRunning Qwen3.8 at 255 tok/s on a single 5090
有团队通过后训练、量化和内核特化技术，让 Qwen3.8-Flash-Next-NVFP4 在单张 RTX 5090 上以 255 tok/s 运行。该模型面向编码和日常任务，推荐硬件为 1x RTX6000，可支持 4 名用户、聚合吞吐 334 tok/sec。
Heat over time
Not enough continuous observations to draw a trend yet.