• Pricing
DashboardContact Sales
  • Blogs
  • Discover
  • Docs
  • Community
  • AgentBeta
⌘K

Over 80,000,000 Cloud Servers Launched

Over 80,000,000 Cloud Servers Launched

Cloud ComputeCloud GPUBare MetalFile SystemObject StorageBlock StorageManaged DatabasesCDNServerlessKubernetesContainer RegistryDirect ConnectLoad Balancers
RegionsAdvanced NetworkControl PanelOperating SystemsUpload ISO
Industry CloudOne-Click DeploymentUse Cases
Browse AppsBecome a Vendor
FAQDevelopers / APIsVultr DocsServer StatusBug BountyPromotionsSolution PartnersStart-Up Programs
Our TeamNewsBrand AssetsReferral ProgramCreator ProgramCareersSLALegalVultr Trust CenterContactYour Privacy ChoicesSubprocessorsAccessibility
Contact SalesSign Up

© Vultr 2026 | VULTR is a registered trademark of The Constant Company, LLC.

Terms of ServiceAUPDMCAPrivacy PolicyCookie Policy
Model Library
Arcee AI
PoolsideReflection AI
Hy4 FlashHy4Hy4 Preview
Tencent Hy3
  1. Docs
  2. More
  3. Model Library
  4. Tencent Hy4
  5. Hy4 Preview

Hy4 Preview

Hy4 preview is a flagship Mixture-of-Experts model designed for advanced reasoning, coding, agentic workflows, and efficient long-context processing. It features 770B total parameters with 49B activated per token, using 78 layers with 6,144 hidden size and 64 attention heads. The architecture combines Gated DeepSeek Sparse Attention with IndexCache, 256 routed experts and 1 shared expert, activating 8 routed experts per token, while iHC expands residual information flow across 4 streams. It includes a native Multi-Token Prediction layer for speculative decoding and supports up to a 1M-token context window.

Deploy with NVIDIA B200Contact Sales

About this model

BlogHF Model Card

Hy4 preview is a flagship Mixture-of-Experts model designed for advanced reasoning, coding, agentic workflows, and efficient long-context processing. It features 770B total parameters with 49B activated per token, using 78 layers with 6,144 hidden size and 64 attention heads. The architecture combines Gated DeepSeek Sparse Attention with IndexCache, 256 routed experts and 1 shared expert, activating 8 routed experts per token, while iHC expands residual information flow across 4 streams. It includes a native Multi-Token Prediction layer for speculative decoding and supports up to a 1M-token context window.

Model Type

MoE LLM

Released

August 28, 2026

License: Apache 2.0

Parameters

770B

Model ID
tencent/Hy4-preview
Capabilities
Text GenerationInstruction FollowingReasoningMathematical ReasoningCode Generation+4

Model specifications and capabilities are published by the model author and reproduced here from HF Model Card (tencent/Hy4-preview). Released under Apache 2.0. Further reading: Blog

Deploy Private Inference Endpoint

Choose your hardware and inference engine to get deployment commands and performance benchmarks tailored to your infrastructure

Hardware

B200

1,440 GB

Engine

vLLM
SGLang

Deployment code

INFERENCE COMMAND
Access Restricted

Enter your email to get access to this content

Powered by Vultr AgentBeta

Requires vllm/vllm-openai:hy4-preview or any standard vLLM image v0.29.0+ to support this model configuration.

Performance under load

23k18k14k9.2k4.6k0163264128256512Total Tokens / secConcurrent requests
Tokens/s (total)

Benchmarks measured by Vultr on B200 · vLLM. Throughput and latency vary with concurrency, input length and engine configuration, so treat these as a comparison baseline, not a service guarantee.

SS

Shivam Sharma

View Profile
View Profile

NVIDIA HGX B200

Unlock Next-Generation AI Power on Vultr Cloud GPU

Deploy NowContact Sales

NVIDIA HGX B200

Starting Price
$3.50
/GPU/hr
Deploy Now

Tech Talk: Running a Self-Improving AI Agent with Hermes on Vultr

From The Blog

Optimize CPU Spend to Accelerate AI: A FinOps Approach to Infrastructure Investment

November 18, 2026

KubeCon + CloudNativeCon 2026: Growing Our Open Source Community in Salt Lake City

November 16, 2026

Optimizing Agentic AI Workflows Through Full-Stack Infrastructure

November 13, 2026

View all posts