Publié le 29/06/26
The fastest way to get this model running locally is via Docker.
Just follow the guidelines provided below.
The system automatically triggers a cloud download for all heavy weights.
The setup file includes an intelligent feature that instantly optimizes all configurations for your hardware profile.
DeepSeek-R1-0528-NVFP4-v2 is a large language model optimized for low‑precision inference on NVIDIA’s Hopper architecture. It leverages NVFP4 data type to achieve higher throughput while maintaining state‑of‑the‑art accuracy. The model features a parameter count of 180 B and was trained on over 5 trillion tokens, enabling robust reasoning across diverse domains. Its inference latency averages 23 ms per token on a single A100‑80GB, making it suitable for real‑time applications. The design incorporates mixture‑of‑experts layers that dynamically route queries to specialized subnetworks, improving both efficiency and scalability. Below is a quick comparison of key technical specifications:
| Parameter Count | 180 B |
| Training Tokens | 5 trillion |
| Inference Latency | 23 ms/token |
| Precision | NVFP4 |
