Deploy Tencent's 295B AI Model: Multi-Node Kubernetes Guide

About this lesson

Deploy Hy3-preview, Tencent Hunyuan's 295B Mixture-of-Experts flagship model across two NVIDIA H100 nodes on Hyperstack Kubernetes. This multi-node LLM deployment tutorial covers LeaderWorkerSet, vLLM tensor parallelism and a full multi-agent code-review pipeline. Hy3-preview has 295B total parameters with only 21B activated per token, a 3.8B MTP speculative decoding layer and a 256K token context window, making it one of the most capable open-source reasoning models available. Because the BF16 checkpoint is ~600GB, a single 8×H100-80GB node isn't enough: this guide shows you exactly how to span it across two nodes using Kubernetes. In this tutorial, you'll learn: Why Hy3-preview requires multi-node deployment (and how to set it up) How to create a 2-node Kubernetes cluster on Hyperstack with H100 NVLink workers How to install the LeaderWorkerSet (LWS) controller How to pre-download 600GB model weights to each node's NVMe storage How to deploy vLLM with TP=16 across both nodes using the LWS manifest How to use Hy3's three reasoning modes: no_think, low, and high How to build a 4-agent code-review pipeline on a single Hy3 deployment Full tutorial on Hyperstack Blog: https://bit.ly/4g1qzCt Get started on Hyperstack: https://bit.ly/4dPt1L9 Found this useful? Like the video and subscribe for more LLM deployment and GPU cloud tutorials! #Hy3preview #Kubernetes #MultiNodeLLM #vLLM #Hyperstack #TencentHunyuan #H100 #LLMDeployment #MixtureOfExperts

Original Description