Tencent Hunyuan Family Overview: From the hy4 Lightweight Line to the A13B MoE Flagship
Tencent Hunyuan Family Overview: From the hy4 Lightweight Line to the A13B MoE Flagship
Tencent Hunyuan is not a single model series on HuggingFace. It is a product matrix: conversational text, document OCR, 3D asset generation, video generation, and image generation each run as their own line, and their repositories live under two different organizations. The most common mistake for a developer meeting Hunyuan for the first time is not model selection — it is finding the wrong organization. Searching Tencent-Hunyuan returns only image generation models and zero text models. This article lays the whole family out using public HuggingFace data: what each line actually does, what the download counts reveal, how to pick a quantization, and whether the hy4 lightweight line is worth deploying.
The 30-second version
The text-line leaders are A13B (the MoE flagship, a clear first in downloads) and hy4 (the 4B lightweight line, the only domestic 4B-class model that ships FP8, GPTQ and AWQ weights from the source). On the non-text side, HunyuanOCR leads with roughly 615,000 monthly downloads, followed by the 3D line. Text models sit under the tencent organization while image models sit under Tencent-Hunyuan — check the organization name before you download anything.
Repository ownership: get the organization right
This is the single highest-frequency mistake in the Hunyuan ecosystem. There are two organizations on HuggingFace:
| Organization | What it holds | Representative repositories |
|---|---|---|
tencent | Text, translation, OCR, 3D, video, image | Hunyuan-4B-Instruct, Hunyuan-A13B-Instruct, HunyuanOCR, Hunyuan3D-2, HunyuanVideo, HunyuanImage-3.0-Instruct |
Tencent-Hunyuan | Image generation only | HunyuanDiT-v1.1 and v1.2, Distilled, ControlNet for Pose, Depth and Canny |
What happens when you search the wrong organization
Searching Tencent-Hunyuan returns nothing but DiT image models. You will conclude that Tencent never open-sourced text models and end up installing a third-party re-upload, where the quantization tier is unknown and the source is unverified. For text models, search Hunyuan directly inside the tencent organization.
Two more names that cause confusion:
- Hunyuan-MT-7B: the
MTstands for machine translation. It is a multilingual translation model, not the successor to the 7B line. - Tencent-Hunyuan-Large: an early Hunyuan repository that is no longer the main line.
The text line: what each size is actually for
| Model | Class | Positioning | Monthly downloads |
|---|---|---|---|
| Hunyuan-0.5B-Instruct | 0.5B dense | Edge devices, classification, tagging | ~2,500 |
| Hunyuan-1.8B-Instruct | 1.8B dense | Lighter edge work, pipeline front-ends | ~2,000 |
| Hunyuan-4B-Instruct (hy4) | 4B dense | Lightweight mainstream, Chinese-first | ~1,000 |
| Hunyuan-7B-Instruct | 7B dense | Mid-size general purpose | ~5,200 |
| Hunyuan-A13B-Instruct | MoE | The family flagship | ~68,700 |
Download counts are a proxy for real community adoption, not an official recommendation. They tell you two things:
- A13B is the de facto Hunyuan flagship — an order of magnitude ahead of the runner-up, which means people are actually running it in production or heavy personal projects.
- hy4's ecosystem is still thin — roughly 1,000 downloads against A13B's 68,700. hy4's value is not community momentum; it is that Tencent ships all three quantization formats for it, which matters when your inference stack has a specific requirement.
Why hy4 is still worth a look despite low downloads
Community download distribution reflects who is running serious projects, not who is better. hy4's differentiation runs on a different axis: the source publishes FP8, GPTQ-Int4 and AWQ-Int4 weights. Most domestic 4B-class models ship only FP8 and leave 4-bit entirely to community converters — hy4 hands the choice to you.
The flagship: how the A13B MoE is structured
A13B is the heaviest open text model Tencent has released: roughly 13B total parameters with far fewer active per token, so inference cost lands closer to a mid-size dense model while the knowledge capacity targets something larger.
Three practical consequences:
- Memory follows total parameters for weights, active parameters for compute. All of the total weight still has to sit in memory; that part of MoE is not free.
- Three official weight sets exist: FP8, GPTQ-Int4, and an official GGUF (
tencent/Hunyuan-A13B-Instruct-GGUF, roughly 540 monthly downloads). Official GGUF releases remain rare among domestic open models. - Engine version matters: a MoE architecture needs llama.cpp or vLLM with Hunyuan support. Older builds fail to load or emit garbage. If it will not start, upgrade the engine before you question the weights.
The quantization matrix: what is official and what is community
| Model | Official FP8 | Official GPTQ-Int4 | Official AWQ-Int4 | Official GGUF | Community GGUF |
|---|---|---|---|---|---|
| hy4 (4B) | Yes | Yes | Yes | No | Several authors |
| A13B | Yes | Yes | No | Yes | A few |
| Hunyuan-7B-Instruct | Yes | No | Yes | No | Several authors |
Community GGUF publishers for hy4 include bartowski, mradermacher, gabriellarson, DevQuasar and Disya, with bartowski's conversion topping the download counts at over 2,000. When picking a community GGUF, favour high download counts, a clearly labeled quantization tier, and a model card that states the conversion source.
Apple Silicon users have a second path in MLX: mlx-community publishes Hunyuan-4B-Instruct-4bit, Hunyuan-A13B-Instruct-4bit and Hunyuan-MT-7B-4bit conversions.
Choosing a quantization in one line
Running a vLLM stack, take the official GPTQ-Int4 or AWQ-Int4. Only llama.cpp or Ollama available, pick the most downloaded community Q4_K_M or Q5_K_M. On a Mac, take MLX 4bit. All three paths work — verify the lowest tier runs end to end, then step up.
The non-text lines: Hunyuan is really a multimodal company
The bulk of Hunyuan's downloads does not come from the text line at all:
| Product line | Representative repository | Monthly downloads |
|---|---|---|
| Document OCR | HunyuanOCR | ~615,000 |
| 3D generation | Hunyuan3D-2 / 2.1 / 2mini | ~319,000 / ~106,000 / ~96,000 |
| Video generation | HunyuanVideo | ~28,700 |
| Image generation | HunyuanImage-3.0-Instruct | ~27,400 |
The toolchains of these lines do not overlap with the text line:
- OCR runs a dedicated document parsing pipeline trained on document layouts, not web text.
- 3D and video run their own diffusion stacks (3D has mesh and texture branches) with no overlap with a language model runtime.
- Image generation is a DiT architecture with different weight formats, quantization approaches, and serving frameworks.
Do not treat Hunyuan as one toolchain
The statement "Hunyuan models run under llama.cpp" only holds for the text line. The 3D and video lines carry their own dependency stacks, often involving point clouds, meshes, and rendering components. If your goal is generating 3D assets, scope that work separately instead of reusing text deployment conclusions.
How to deploy Hunyuan text models
| Situation | Recommended path |
|---|---|
| Single consumer GPU | Official GPTQ-Int4 or AWQ-Int4, served through vLLM |
| CPU-only or Mac | Community GGUF or MLX with llama.cpp / Ollama |
| Internal team inference service | Official FP8, keeping quality headroom |
| Fine-tuning research | Start from official FP8 or BF16 weights, not a 4-bit quantized file |
The generic flow never changes: pull from HuggingFace, convert the weights if needed, load them in your inference framework, expose an OpenAI-compatible endpoint, then point your existing client at a new base URL.
The EXAIV cloud catalog is the source of truth
At the time of writing, the EXAIV cloud catalog does not carry the Hunyuan family (the domestic models on sale are Qwen, DeepSeek, Kimi and GLM). Hunyuan means deploying the weights in your own environment: official quantizations with vLLM, or community GGUF and MLX with llama.cpp and Ollama. The main platform catalog lists its own IDs such as hy3, hy4, hy3-preview and hy4-preview — always check the live catalog and never treat a HuggingFace weight name as an API model ID.
Selection cheat sheet
| Your situation | Pick Hunyuan |
|---|---|
| Chinese-first assistant, edge deployment | hy4 (4B) |
| Ultra-light edge work, classification | Hunyuan-0.5B-Instruct |
| You need 4-bit maintained by the source | hy4 (GPTQ-Int4 / AWQ-Int4) |
| Highest open capability, MoE | A13B |
| Multilingual translation | Hunyuan-MT-7B |
| Document parsing, layout recovery | HunyuanOCR |
| 3D asset generation | Hunyuan3D-2 series |
| Video generation | HunyuanVideo |
| Image generation | HunyuanImage-3.0-Instruct |
FAQ
Which HuggingFace organization holds the Hunyuan text models?
How do I choose between hy4 and A13B?
Is there an official GGUF for A13B?
Is there an official context window number for Hunyuan models?
Why do Hunyuan text models download far less than OCR and 3D?
Can I call Hunyuan through an API?
What is Hunyuan-MT-7B?
Related reading
- hy4 model page — specification, quantization matrix and memory math
- hy4-preview channel notes — what the 1M context channel is for
- Hunyuan hy4 and A13B local deployment — deployment detail for the two main text lines
- hy4 model selection guide — how hy3, hy4, the preview builds and A13B divide the work
- hy4 GGUF local deployment — quantization tiers and VRAM math
- Qwen vs Hunyuan — ecosystem activity versus quantization completeness
- DeepSeek V4 family overview — the other thoroughly open domestic family
- Model directory — models callable on EXAIV today
Sources
Download figures are HuggingFace public 30-day counters and move over time; the numbers here were sampled in November 2026. Licences and parameter specifications follow the live official model cards.