DeepSeek-V4.1-Flash-1017-NVFP4: 552B MoE model, NVFP4 checkpoint compressed to 400GB

This is the NVFP4 quantized version of DeepSeek-V4.1-Flash-1017. Using NVIDIA Model Optimizer, the routing MoE experts were converted to W4A4 precision, and the checkpoint size is approximately 400 GiB. It is designed to reduce capacity compared to the original model while maintaining inference performance.

It is based on a multimodal MoE architecture that supports a 1 million token context. It accepts text and image inputs and outputs text, activating only 8B parameters during prefill and 16B during decoding out of 552B parameters. It also includes 196B Engram conditional memory.

MXFP4 Engram support is required for vLLM deployment. It cannot be loaded with the standard vLLM image; a build with this feature added is required. It is optimized for NVIDIA Blackwell GPUs, particularly in GB300 environments.

It shows performance similar to the original MXFP4 model in benchmarks such as GPQA Diamond 92.93% and MMMU-Pro 78.73%. It is released under the MIT license for commercial and non-commercial deployments, including AI agents, chatbots, and RAG systems.