From NVIDIA: 2025-01-30 18:30:58

DeepSeek-R1 is a reasoning model that performs multiple inference passes to generate the best answer, demonstrating test-time scaling. It delivers leading accuracy for tasks like reasoning, math, and language understanding, while also offering high inference efficiency. The 671-billion-parameter model is available as a NVIDIA NIM microservice preview for developers to experiment with.

DeepSeek-R1 is a mixture-of-experts model with 671 billion parameters, supporting a large input context length and extreme number of experts per layer. Real-time answers require high compute performance, with software optimizations enabling the model to run at up to 3,872 tokens per second on a single server with eight H200 GPUs.

Developers can access the DeepSeek-R1 NIM microservice on build.nvidia.com to experience its capabilities. The NVIDIA NIM platform allows for easy deployment and high efficiency for agentic AI systems.



Read more at NVIDIA: DeepSeek-R1 Now Live With NVIDIA NIM