nvidia/Qwen3.8-Flash-Next-FP4-Local — new model trending #30 on Hugging Face
NVIDIA published an FP4-quantized Qwen3.8-Flash-Next checkpoint for local Blackwell inference, occupying about 96 GiB.
NVIDIA published Qwen3.8-Flash-Next-FP4-Local, an FP4 quantization of Alibaba's multimodal Qwen3.8-Flash-Next, produced with Model Optimizer v0.47.0 and quantization-aware distillation. The checkpoint is about 96 GiB, with 125B total parameters and 6B activated, plus 51B n-gram embeddings and 4B MTP. It accepts text, image, and video, with 262K native context extensible to 1 million tokens. NVIDIA validated it on an RTX PRO 6000 Blackwell and DGX Spark using vLLM, but the new NVFP4 scale format still requires NVIDIA containers because upstream vLLM support has not landed.
- FP4 checkpoint of Alibaba Qwen3.8-Flash-Next, about 96 GiB on disk.
- 125B total parameters, 6B activated, plus 51B n-gram embeddings.
- Validated on RTX PRO 6000 Blackwell and DGX Spark with vLLM.
- Native 262K context, extensible to 1 million tokens.
- New NVFP4 scale format needs NVIDIA containers; upstream vLLM support pending.
- Products
- vLLMNVIDIA Model Optimizer
From the original article
8 min read at huggingface.coimage-text-to-text · 771 downloads · 151 likes · license other
By nvidia · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.