nerv-dynamics/Qwen3.8-Flash-Next-S2

Qwen3.8-Flash-Next for S2

Qwen/Qwen3.8-Flash-Next quantised to NVFP4 for S2, the inference engine of Nerv Linux, with PixelML's DFlash drafter for speculative decoding.

file	what it is
Qwen3.8-Flash-Next-s2v1p.s2w	the model: NVFP4 weights in S2's format, with its tokenizer and chat format
Qwen3.8-Flash-Next-dflash-pixelml.s2w	the drafter, in S2's format

These files are S2's own format (.s2w), not safetensors: they run on S2 and nothing else.

Run it

On Nerv:

dnf install nerv-s2
s2 pull qwen3.8-flash-next   # both files, checked against the SHA-256 the package records
s2                           # the dashboard: n, then Enter

The first run prepares the model for the machine's GPU, once. S2 serves this model's sparse attention as its exact dense equivalent, which covers 2051 positions a session.

Licence

Built with Qwen. The weights are quantised from Qwen3.8-Flash-Next and are under the Qwen Community License 1.0, whose text is LICENSE; read its conditions before commercial use. The drafter is PixelML's, repacked: it was trained on outputs of Qwen3.8-Flash-Next served through NVIDIA's NVFP4 build of it, so the Qwen Community License 1.0 and the NVIDIA Open Model License apply to it. Licensed by NVIDIA Corporation under the NVIDIA Open Model License.