NVIDIA Shrinks DeepSeek V4.1 Flash by 75 GiB With Almost No Quality Loss

NVIDIA has published a quantized build of DeepSeek-V4.1-Flash-1017 that reduces the checkpoint from about 475 GiB to 400 GiB, a decrease of roughly 16%. Across six reported reasoning, coding, vision, instruction-following, long-context, and agent benchmarks, scores move by no more than 0.70 points. The NVFP4 model card targets Blackwell GPUs through vLLM and carries an MIT license for commercial and noncommercial use.

Only part of the model runs per token

The base release is a natively multimodal Mixture-of-Experts model with a Causal Encoder-Decoder backbone, Compressed Sparse Attention 2, and a context window of up to one million tokens. DeepSeek designed it to process text and images while limiting the attention work required for long prompts.

Its backbone contains 552 billion parameters, but the routing system activates about 8 billion while processing a prompt, known as prefill, and 16 billion while generating tokens, known as decoding. A separate 196 billion-parameter Engram module provides conditional memory through lookup tables.

Where NVIDIA spends the 4 bits

NVIDIA produced the checkpoint through post-training quantization with Model Optimizer . The standard routed experts use NVFP4 weights and activations, described as W4A4, with one scale for each group of 16 values. The conversion covers the w1 , w2 , and w3 projections across 384 experts and 40 layers.

The Engram lookup tables move from FP8 to MXFP4. Attention layers, shared experts, vision components, and multi-token prediction components retain their source precision. This selective approach concentrates compression on the model’s largest routed tensors while preserving higher precision in components that may be more sensitive to quantization.