Skip to content
English

Switch theme

// Tencent Hunyuan / Hy · Image generation

HunyuanImage-3.0-Instruct: VRAM & local deployment

80B total · 13B active (MoE) parameters. About 53.9 GB at 4-bit and 177 GB at BF16.

released 2025-09-25licence: Custom licence – read itmemory figures are estimates

01Specs

PublisherTencent · tencent
Family / seriesTencent Hunyuan / Hy · HunyuanImage-3.0
TypeImage generation (image-to-image)
Parameters80B total · 13B active (MoE) Source: official model card“80B total (13B active)”
Released weightsBF16 · 169 GBdownload
Licenceother · Custom licence – read itThe publisher uses its own licence; we have not classified it. Read the licence text.
Released2025-09-25 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Image models: memory depends on resolution, batch and which text encoders are kept on the GPU; Diffusers / ComfyUI can offload parts to CPU. The weight size below is the main transformer only.

PrecisionWeightsWeights + overheadSmallest GPUMac
BF16 / FP16 (original)160 GB 177 GB 4 × RTX PRO 6000 (384GB)Mac 512GB
FP8 / INT8 / Q8_085.0 GB 94.0 GB H200 141GBMac 256GB
4-bit (Q4_K_M / AWQ / GPTQ-Int4)48.5 GB 53.9 GB 2 × RTX 5090 (64GB)Mac 128GB

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

2 official repo(s) from tencent.

05Other Tencent Hunyuan / Hy models

All 19 →
ModelParams4-bit totalReleased
Hy-Memory-30B-A3B30.1B / 3B21.4 GB2026-10-15
Hy4-preview770B / 49B515 GB2026-08-27
Hy3295B / 21B200 GB2026-07-02
Hy-MT2-1.8B2.04B2.4 GB2026-05-11
Hy-MT2-7B8.03B6.9 GB2026-05-11
Hy-MT2-30B-A3B30.1B / 3B21.4 GB2026-05-11
Hy3-preview295B / 21B200 GB2026-04-13
HY-MT1.5-1.8B2.04B2.4 GB2025-12-25

06Sources

Data checked 2026-02-03. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)