Tencent Hunyuan · 28 Aug 2026

Hy4 preview

Name: Hunyuan Hy4 preview
Author: Tencent Hunyuan

Tencent Hy4 — also called Hy4 preview and Hunyuan 4 — is Tencent Hunyuan’s open-weight text model. 770B MoE, 49B active, 1M context. This page is the LLM, not a projector or the H2FLY aircraft.

Download official weights

Official Hy4 page

770B: total parameters
49B: active parameters
1M-token: context
Apache 2.0: license

What is Tencent Hy4?

Hy4 preview (hy4-preview, Hunyuan 4, hunyuan hy4) is the current public Hy4 model. Three questions cover the observed searches.

Hy4 identityWhat is Hy4 preview?Official name, whether it has shipped, and whether Hunyuan 4 is the same Hy4 model.ModalityIs Hy4 multimodal?This preview is text-only. It cannot natively see images, charts, or video.Generation stepHy4 vs Hy3What changed from Hunyuan 3: scale, context, and whether you should switch.

All Overview

Hy4 preview specs

Numbers below are from the public config and model card on 28 August 2026. They describe the checkpoint, not a community rerun.

770B MoE, 49B active

78 layers, Top-8 of 256 routed experts plus one shared expert. ~780B if you count the MTP module.

Architecture

1M-token context

config.json sets 1,048,576 tokens. That is native in config. Quality at 1M is still an open measurement.

Context window

Apache 2.0 weights

BF16 and FP8 instruct checkpoints on Hugging Face: tencent/Hy4-preview and tencent/Hy4-preview-FP8.

Weights & APIs

Text in, text out

Hy4 preview cannot see. Later Hunyuan 4 versions may add vision; this checkpoint does not.

Modality

Gated DSA

Sparse attention so each token does not attend to the full history. Published speed-ups are from a 30B paper, not a 780B curve.

Architecture

Hy4 vs Hy3 scale

Context config goes from about 262K to 1M. Agent scores jump on vendor charts. API list price is also several times Hy3.

Hy4 vs Hy3

Launch-day card

Official label: Hy4 preview
Also called: Tencent Hy4 · Hunyuan 4 · Hunyuan 4.0
Released: 28 August 2026
Architecture: MoE, 78 layers, Top-8 of 256 experts
Parameters: 770B backbone · ~49B active / token
Context: 1,048,576 tokens (config)
Hidden / heads: 6,144 hidden · 64 attn · 8 KV
License: Apache 2.0
Modality: Text input and text output
Hugging Face: tencent/Hy4-preview · tencent/Hy4-preview-FP8

All Overview Hy4 vs Hy3

How to use Hy4 preview

Try it in a Tencent product, call an API, or download weights. A normal PC cannot host this model.

Try Tencent Hy4

Product trials and chat entry points, without standing up a cluster.

Download the Hy4 model

Official instruct weights on Hugging Face. FP8 is about 814GB; BF16 is about 1.56TB.

Hugging Face

API or self-host

TokenHub id hy4, OpenRouter tencent/hy4-preview. Official vLLM examples use 16×B200 or 8×B300.

Hy4 API

OpenRouter All weights All comparisons

Hy4 vs Hy3 at a glance

Hy4 is a generation step, not just a rename. The largest visible jump is the context configuration, alongside a much larger MoE backbone.

DimensionHy3Hy4 preview

Context config~262K tokens1,048,576 tokens

BackbonePrior Hunyuan generation770B MoE · ~49B active

AttentionDense generation baselineGated DSA

ModalityTextText only

LicenseModel-card specificApache 2.0

FAQ

Short answers dated 28 August 2026. This is Tencent Hunyuan Hy4 preview, not H2FLY HY4 or a projector.

What is Hy4 preview?

Hy4 preview is Tencent Hunyuan’s open-weight text MoE, also called Tencent Hy4, Hunyuan 4, and hunyuan hy4. Released 28 August 2026: about 770B backbone, ~49B active per token, Apache 2.0 instruct weights.

Is Hy4 multimodal?

No. This preview is text-only. Tencent has said later versions should add vision. WorkBuddy demos that draw UIs are code-and-tool output, not image input.

Hy4 vs Hy3 — should I switch?

It is a generation step, not a rename. Context config goes from about 262K to 1M, vendor agent scores jump, and list price is several times Hy3. Read the comparison before you cut over.

Is this the H2FLY HY4 aircraft?

No. Searches for hy4 also hit a hydrogen aircraft, AL-KO hardware, and projector SKUs. This site is Tencent Hy4, the LLM.

Can I run the Hy4 model locally?

Not on a normal PC. FP8 weights are about 814GB. Official vLLM examples use 16×B200 or 8×B300. There was no mature GGUF pack on launch day. Try it in AI Studio or via an API.

Hunyuan Hy4 preview