NVIDIA cuts DeepSeek Flash download by 75 GiB, reports small benchmark shifts

NVIDIA has released a smaller downloadable version of DeepSeek’s Flash model, cutting its saved model files from about 475 GiB to 400 GiB. Published on October 30, 2026, the checkpoint offers a narrower change than a new model generation: less storage, with scores staying within 0.70 percentage points of the source in NVIDIA’s six reported benchmark comparisons.

The release, DeepSeek-V4.1-Flash-1017-NVFP4, is NVIDIA’s quantized version of DeepSeek AI’s existing model—not a base model NVIDIA developed or owns. It carries the MIT license for commercial and noncommercial use. It accepts text and images and produces text, with a context window of up to one million tokens—the material it can work with at once.

Different parts get different precision

Quantization changes the numerical precision used to represent a model. Using its Model Optimizer software, NVIDIA converted the routed experts—specialized parts selected by the model—from MXFP4 to NVFP4. Their weights and activations use four-bit precision, with a scaling group of 16 values. The conversion covers 384 experts across 40 layers.

NVIDIA also reduced the precision of the Engram lookup tables, a conditional-memory component, from FP8 to MXFP4. Attention, shared experts, vision and multi-token prediction components keep their source precision. The download is therefore a mixed-precision package, not a uniformly four-bit model.

Small score changes across six tests

NVIDIA’s evaluation spans scientific reasoning, long-context recall, scientific coding, instruction following, visual understanding and terminal-based agent tasks. Three scores increased and three decreased. The largest drop was on Terminal-Bench 2.1, from 81.81% to 81.11%. GPQA Diamond rose from 92.77% to 92.93%, while the visual test MMMU-Pro rose from 78.38% to 78.73%.

NVIDIA reports 16 repeats for GPQA, eight runs averaged for SciCode, and two runs averaged for Terminal-Bench. The terminal evaluation used a separate agent setup with eight trajectories per task.

The runtime requirement survives the compression

The release targets NVIDIA Blackwell hardware. NVIDIA’s reference deployment used four GB300 GPUs and vLLM, software that loads models and serves requests. Its deployment instructions require a build containing MXFP4 Engram support from vLLM pull request 61530; stock vLLM images cannot load this checkpoint’s lookup tables.

What NVIDIA validated

• Text and image inputs were validated with the published serving command, which enables CPU offloading for Engram.
• Multi-token prediction, a speculative-decoding feature, was not exercised in that setup.
• SGLang, another model-serving runtime, has not been validated with this checkpoint.