Qwen3.8-Flash-Next NVFP4 (109 GB, single-GPU vLLM)

A 109 GB NVFP4 quantization of the 335 GB Qwen/Qwen3.8-Flash-Next that runs through vLLM on a single 121 GB GPU. The 51.2B-parameter n-gram (PLE) table is quantized to 4-bit; a small vLLM loader patch (linked below) makes it loadable.

Requires the loader patches + serving instructions here: https://github.com/starkweatherdigital/qwen3-8-flash-next-nvfp4 (prebuilt image: docker.io/jstarkg/vllm-gb10-flashnext:0.28-sm121-r3, serve with VLLM_PLE_NVFP4=1).

Measured on DGX Spark GB10: 24.6 tok/s single-stream with MTP speculative decoding (80% acceptance), 16.8 tok/s without. Full recipe, verification, provenance notes, and benchmarks in the GitHub repo.

License: Qwen Community License 1.0 (inherited; see LICENSE).

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for starkweatherdigital/qwen3.8-flash-next-nvfp4

Quantized
(85)
this model