NVIDIA releases 96 GiB NVFP4 build of Qwen3.8-Flash-Next for single-GPU Blackwell and DGX Spark

NVIDIA has published Qwen3.8-Flash-Next-FP4-Local, a roughly 96 GiB NVFP4 quantization of Alibaba's Qwen3.8-Flash-Next, on Hugging Face. NVIDIA says the checkpoint runs on a single NVIDIA RTX PRO 6000 Blackwell GPU or a DGX Spark. The Hugging Face listing gives a release date of 10/05/2026.

This is a quantized derivative, not a new base model. NVIDIA states the model is "not owned or developed by NVIDIA" and was built from Alibaba's Qwen3.8-Flash-Next.

Architecture and specs

According to the model card, the underlying model has:

• Parameters: 125B total, 6B activated, plus 51B of n-gram embeddings and 4B for multi-token prediction (MTP)
• Architecture: Mixture-of-Experts with hybrid attention (Gated DeltaNet and Qwen Sparse Attention), gated residual streams, n-gram embeddings and a vision encoder
• Inputs: text, image and video; output is text
• Context window: 262K tokens natively, extensible to 1,000,000 tokens

Pricing is not applicable, since this is a downloadable checkpoint. The training cutoff date is not disclosed.

Hardware validation

NVIDIA reports validation with vLLM on two systems:

• RTX PRO 6000 Blackwell (single GPU): 850K tokens of FP8 KV cache, with PLE (n-gram embedding table) offload
• DGX Spark: 1M tokens of FP8 KV cache with PLE offload; it can also run without offload

The card lists only Blackwell as a compatible microarchitecture, Linux as the preferred OS, and vLLM as the only supported runtime.

Accuracy results

NVIDIA's own evaluations, against the BF16 baseline of Qwen3.8-Flash-Next:

Precision	GPQA Diamond	ToolEvalBench
BF16	92.2%	89.4%
FP4-Local	91.5%	88.7%

That is a 0.7-point drop on both benchmarks. These figures come from NVIDIA's model card and have not been independently verified.

How it was quantized

NVIDIA used Model Optimizer (nvidia-modelopt v0.47.0) with quantization-aware distillation (QAD). The BF16 teacher supplies token-level probability and hidden-state targets. The loss combines total-variation probability matching, hidden-state matching, and separate losses for end-of-response and end-of-thinking behavior.

Training ran 2,500 trunk updates followed by 1,500 joint-refinement updates. It used 500 million tokens repeated for 3 epochs, with prompts drawn from the Nemotron post-training and pretraining datasets.

Precision varies by component:

• Decoder routed experts: NVFP4
• Decoder attention: MXFP8
• Vision MLP and merger linears: NVFP4
• Vision attention linears: MXFP8
• Normalization parameters and other vision tensors: BF16 or FP32

Deployment caveat

The NVFP4 scale factors use a new lossless compression format, and upstream vLLM support has not landed. Users must run one of the containers NVIDIA provides.

Licensing

Use is governed by the NVIDIA Open Model Agreement, and the base model remains under the Qwen Community License 1.0. NVIDIA states the model is cleared for commercial and non-commercial use.

What this means

The notable point is the memory footprint. A 125B-parameter multimodal MoE with up to 1M-token context now fits on a single workstation GPU or desktop-class DGX Spark, which previously meant multi-GPU servers for models this size. The 6B active-parameter count keeps per-token compute low, and PLE offload handles the memory-heavy embedding tables.

The tradeoffs are clear. The checkpoint is Blackwell-only and depends on custom containers until vLLM merges support, which limits portability. The accuracy figures are NVIDIA's own and cover only two benchmarks, so long-context and vision quality at FP4 remain untested publicly. Even so, the sub-1-point loss on both benchmarks suggests QAD is a viable route to local deployment of large MoE models.