Four GB300. That is how many GPUs stand behind NVIDIA's numbers for the new DeepSeek V4.1 Flash. The version is quantized to NVFP4 and weighs about 400 GiB instead of the original's roughly 475. It sounds like something you could fit somewhere by now. Not so fast: the card lists only Blackwell.
Fact
The facts: on 30 October 2026 NVIDIA published DeepSeek-V4.1-Flash-1017-NVFP4 on Hugging Face, the quantized version of DeepSeek AI's model, made with Model Optimizer. Per NVIDIA, the model is not owned or developed by NVIDIA, and the license is MIT. The backbone is 552 billion parameters in total, 8 billion activated during prefill and 16 billion during decode, plus 196 billion of Engram conditional memory; context runs up to 1 million tokens. The checkpoint is about 400 GiB (source: about 475 GiB) in 48 safetensors shards. The supported runtime is vLLM, but the Engram lookup tables are stored as MXFP4 and stock vLLM images cannot load them: you need a build with pull request vllm-project/vllm#61530. SGLang has not been validated with this checkpoint. The supported microarchitecture is NVIDIA Blackwell; the results come from four GB300 with vllm/vllm-openai:v0.31.0 and that pull request's changes.
Accuracy, per NVIDIA, barely moves. GPQA Diamond: 92.77 for the MXFP4 source and 92.93 for NVFP4. AA-LCR: 80.41 and 80.19. Terminal-Bench 2.1: 81.81 and 81.11. Two of the three drop, one rises. The numbers are NVIDIA's, measured with vLLM, and the page has no outside check.
The lighter version still weighs about 400 GiB.
The more interesting part is in the card's fine print. The expert weights passed conversion without loss. The Engram conversion, however, is lossy. And one more sentence from the same card: lossless weight conversion does not imply identical inference outputs.
What was not tested is written down too. Speculative decoding (MTP) was not exercised. SGLang has not been validated. The card lists only Blackwell, so how it runs elsewhere, we don't know. We also don't know when patch #61530 lands in vLLM.
If you plan to run it, get a vLLM build with the patch first. The stock image will not load the tables. Then validate the final package with your own runtime, as NVIDIA itself says.
Official primary sources
NVIDIA - DeepSeek-V4.1-Flash-1017-NVFP4 model card (Hugging Face), 30.10.2026The visual is generated code-art, no third-party images.