AI BriefingNVIDIA Unveils ModelExpress
·2026.07.25 17:24
Key point
NVIDIA has unveiled ModelExpress, which dramatically speeds up model loading by leveraging GPU-to-GPU RDMA.
Details
NVIDIA succeeded in reducing the startup time of DeepSeek-V4 Pro from 8 minutes to under 2 minutes through ModelExpress (MX). This technology operates as a weight distribution and cache management service for NVIDIA Dynamo.
The key technical features are as follows:
- Leveraging GPU-to-GPU RDMA: Maximizes data transfer efficiency by using the fastest path for moving weights into GPU memory.
- NIXL-based direct fetching: Inference workers fetch updated weights directly from other GPUs via NIXL, without going through a centralized broadcast.
- Path optimization: Improves the speed of both inference and RL (reinforcement learning) post-training by decoupling weight movement from the critical path.
- Kernel cache reuse: MX reuses kernel caches to reduce resource waste and increase efficiency.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Stanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster InferenceDeepSeek Releases DSec: Elastic Sandbox Infrastructure for Large-Scale Agentic Training