Qwen3.8-Flash-Next

Name: OpenWeights Terminal measured pairs
Creator: Dvidia
License: https://owterminal.com/cite

DGX Spark · NVFP4 · TensorFold

Decode · tok/s

Prefill

—

tok/100W

13.3

在单台 DGX Spark 上用 TensorFold 跑 Qwen3.8-Flash-Next，试着把 Vontra MLX 4-bit模型换成 NVIDIA NVFP4。结果 decode 仅仅32 tok/s。

First-person swap on one Spark. Rank is the quoted 32 decode. Weight file 123.57 GiB.

Officialclaimed