Tencent Hunyuan
The Tencent Hunyuan model family from a local deployment angle: the hy4 lightweight line, the Hunyuan-A13B MoE flagship, plus the OCR, 3D, video and image lines, with weight sources, quantization matrix and selection advice.
Tencent Hunyuan
On HuggingFace, Tencent Hunyuan is a family of parallel product lines: conversational text, document OCR, 3D generation, video generation, and image generation each run separately, with weights split across the tencent and Tencent-Hunyuan organizations. This page maps the family from a local deployment angle: where the weights come from, what quantizations exist, and which member fits which job.
The EXAIV cloud catalog is the source of truth
Whether Hunyuan appears in the EXAIV cloud catalog depends on the live model directory. This page covers the self-hosted route: official quantizations with vLLM, or community GGUF with llama.cpp and Ollama. Never treat a HuggingFace weight name as an API ID.
The text line
| Model | Parameters | Structure | Positioning | Local deployment floor |
|---|---|---|---|---|
| Hunyuan-0.5B-Instruct | 0.5B | Dense | Edge, classification and tagging | Very low |
| Hunyuan-1.8B-Instruct | 1.8B | Dense | Edge assistants, pipeline front-ends | Very low |
| hy4 | 4.22B | Dense | Lightweight mainstream, Chinese-first | From a 6GB GPU |
| Hunyuan-7B-Instruct | 7.5B | Dense | Mid-size general purpose | From an 8GB GPU |
| Hunyuan-A13B-Instruct | 80.39B | MoE, 64 experts | The family flagship | Server-class multi-GPU |
| Hunyuan-MT-7B | 7B | Dense | Multilingual translation | From an 8GB GPU |
Why hy4 is the default starting point for this family
It is the only 4B-class model that ships FP8, GPTQ-Int4 and AWQ-Int4 official weights together, its 4-bit weights are only 2.6GB, and its native context is 262144. Few domestic models clear all three bars at once: low hardware floor, real quantization choice, and a long window.
The non-text lines
| Product line | Representative repository | Toolchain |
|---|---|---|
| Document OCR | HunyuanOCR | A dedicated layout parsing pipeline |
| 3D generation | Hunyuan3D-2 / 2.1 / 2mini | A 3D diffusion stack across mesh and texture |
| Video generation | HunyuanVideo | A video diffusion stack |
| Image generation | HunyuanImage-3.0-Instruct | A DiT architecture |
These stacks share nothing with the text line, so llama.cpp deployment conclusions do not transfer to them.
Where the weights live
| Repository | Contents |
|---|---|
| tencent/Hunyuan-4B-Instruct | Original hy4 BF16 weights |
| tencent/Hunyuan-4B-Instruct-FP8 | Official hy4 FP8 |
| tencent/Hunyuan-4B-Instruct-GPTQ-Int4 | Official hy4 GPTQ 4-bit |
| tencent/Hunyuan-4B-Instruct-AWQ-Int4 | Official hy4 AWQ 4-bit |
| tencent/Hunyuan-A13B-Instruct | Original A13B weights |
| tencent/Hunyuan-A13B-Instruct-FP8 | Official A13B FP8 |
| tencent/Hunyuan-A13B-Instruct-GGUF | Official A13B GGUF, which is rare |
| bartowski/tencent_Hunyuan-4B-Instruct-GGUF | Community hy4 GGUF with the widest tier coverage |
| mlx-community/Hunyuan-4B-Instruct-4bit | hy4 for Apple Silicon |
Selection cheat sheet
| Your situation | Take |
|---|---|
| Chinese assistant, edge, internal network | hy4 |
| Official 4-bit weights you can maintain | hy4 (GPTQ-Int4 / AWQ-Int4) |
| Classification, tagging, routing | Hunyuan-0.5B-Instruct |
| Highest open capability, open MoE | A13B, on multiple GPUs |
| Multilingual translation | Hunyuan-MT-7B |
| Document parsing and layout recovery | HunyuanOCR |
| 3D asset generation | The Hunyuan3D-2 series |
Where to start
- Read hy4 GGUF local deployment and work out your memory budget.
- Read hy4 model selection guide to confirm hy4 is the tier you want.
- For the full family map, see Tencent Hunyuan family overview.
- To wire it into your tools, see Running hy4 with Ollama.
Related reading
- hy4 model page — specification, quantization matrix, and memory math
- Tencent Hunyuan family overview — the full family and its download structure
- Hunyuan A13B MoE in practice — the deployment ledger for an 80B MoE
- Qwen vs Hunyuan — ecosystem comparison between the two families
Ready to build with Hunyuan?
One API Key, every model. Transparent pay-as-you-go pricing with credits that never expire.
Create free account