NVIDIA Drops DeepSeek-V4.1 Flash NVFP4

This is a practical inference optimization rather than a new model, but the modest quality loss makes the tradeoff compelling for teams operating large MoE checkpoints.

• –Coding scores fall only 0.70 points on Terminal-Bench 2.1 and 0.26 points on SciCode.
• –GPQA Diamond slightly improves from 92.77 to 92.93, while AA-LCR dips from 80.41 to 80.19.
• –NVFP4 converts routed experts to W4A4 precision, cutting the checkpoint by about 81 GB while retaining the model’s one-million-token context and multimodal inputs.
• –Deployment currently requires NVIDIA Blackwell hardware and a vLLM build with MXFP4 Engram support; SGLang has not been validated.
• –The open MIT license lowers commercial adoption friction, but the four-GB300 benchmark setup means most teams still need to validate real-world serving costs and throughput.