How to Choose hy4: The Four Hunyuan Lightweight Sizes and Where A13B Fits
How to Choose hy4: The Four Hunyuan Lightweight Sizes and Where A13B Fits
Hunyuan model names make selection surprisingly easy to get wrong. hy3, hy4, hy3-preview, hy4-preview and Hunyuan-A13B-Instruct look like four close relatives, but they actually span different size tiers, different generations, and different release channels. On top of that, the platform-side model IDs do not map one-to-one onto HuggingFace weight names, so selection becomes a two-step job: first understand the naming, then understand the workload. This article lays out every usable Hunyuan size at once — positioning, official quantizations, VRAM floor, and the scenarios where each one is the right answer.
The 30-second version
Edge and embedded devices take hy4 (4B). When memory is tight enough for a 2B-class model, use 1.8B or 0.5B. When you want a higher capability ceiling, take A13B (MoE). When you need 4-bit weights maintained by the source, take hy4 (GPTQ-Int4 / AWQ-Int4). For translation, take Hunyuan-MT-7B. The preview suffix marks a platform channel, not a worse official release.
All five sizes side by side
| Size | Weight shape | Typical use | Approx. footprint at 4-bit | Community activity |
|---|---|---|---|---|
| Hunyuan-0.5B-Instruct | 0.5B dense | Classification, tagging, routing | under 1GB | ~2,500 / month |
| Hunyuan-1.8B-Instruct | 1.8B dense | Edge assistants, pipeline front-ends | 1–2GB | ~2,000 / month |
| Hunyuan-4B-Instruct (hy4) | 4B dense | Lightweight mainstream, Chinese-first | 3–4GB | ~1,000 / month |
| Hunyuan-7B-Instruct | 7B dense | Mid-size general capability | 5–6GB | ~5,200 / month |
| Hunyuan-A13B-Instruct | MoE | The family flagship | 8GB and up | ~68,700 / month |
Footprints are rough 4-bit weight estimates; real usage adds the KV cache and runtime overhead. The exact math is covered in hy4 GGUF local deployment.
Platform IDs and weight names are two naming layers
The platform catalog exposes service-side IDs such as hy3, hy4, hy3-preview and hy4-preview. HuggingFace weights are named Hunyuan-4B-Instruct, Hunyuan-4B-Pretrain, with -FP8, -GPTQ-Int4 and -AWQ-Int4 quantization repositories alongside. Use the latter for local deployment and the former for API calls. Mixing the two is the most common way to waste an afternoon.
hy3 versus hy4: two independent lightweight lines
hy3 maps to the Hunyuan-7B tier, hy4 maps to the Hunyuan-4B tier. The relationship is not a small and large version of one model — they are two separate lightweight lines:
| Dimension | hy3 | hy4 |
|---|---|---|
| Parameters | 7B dense | 4B dense |
| Memory at 4-bit | 5–6GB | 3–4GB |
| Platform context label | 256K | currently unlabeled |
| Official quantizations | FP8 / AWQ-Int4 | FP8 / GPTQ-Int4 / AWQ-Int4 |
| Community GGUF | several authors | several authors, bartowski leading |
Pick based on how much headroom your card has. If a 6GB card runs hy3 without fighting your training job for memory, hy3 gives you a higher ceiling. If the card is a 4GB-class part or shared with another service, hy4 is the realistic option.
Half the parameters is not half the capability
The 4B versus 7B gap shows up mostly in long-form reasoning, complex instruction following, and knowledge coverage. If your task is structured extraction with a fixed output format, hy4 often performs closer to the size difference suggests. If the task is long-form writing or multi-step reasoning, the 7B gain is real.
What the preview suffix means
The platform lists hy3-preview and hy4-preview as separate entries. They are not watered-down official builds — they are different release channels:
| ID | Platform context label | Fits |
|---|---|---|
| hy3-preview | 256K | A stable channel that needs a long context |
| hy4-preview | 1,048,576 (1M) | Very long context: whole repositories, long meeting transcripts, batches of large PDFs |
The 1M context is hy4-preview's most distinctive label. Two caveats:
- A long window does not mean long context works well. Retrieval accuracy, attention dilution, and inference cost all need separate evaluation at 1M. A big window is only a ceiling.
- Labels change. Context lengths and billing factors follow the live catalog, so the numbers here reflect the state at sampling time only.
The right way to read a preview ID
Treat it as "this channel advertises an extra-long context", not as "this build is experimental and unreliable". Whether it helps depends on whether your workload genuinely requires megabytes of material in one prompt. If it does not, the regular hy4 channel is usually the calmer choice.
Choosing between hy4's three official quantizations
hy4 is the only Hunyuan model that ships all three quantization formats at the 4B class:
| Quantization | Quality loss | Fits |
|---|---|---|
| FP8 | Almost none | vLLM / SGLang, teams standardizing on one stack |
| GPTQ-Int4 | Small | AutoGPTQ ecosystem, vLLM |
| AWQ-Int4 | Small, activation aware | AutoAWQ ecosystem, vLLM |
GPTQ and AWQ are different 4-bit algorithms: GPTQ applies per-layer error compensation, while AWQ protects important channels using activation statistics. Results often differ on the same model, and the point of publishing both is that you can measure and choose.
The community GGUF route is equally well served: bartowski, mradermacher, gabriellarson, DevQuasar and Disya all publish hy4 conversions, with bartowski's the most downloaded at over 2,000. Apple Silicon users can also take the mlx-community 4bit build, or the taobao-mnn MNN release for edge deployment.
Do not fine-tune from a 4-bit quantized file
Fine-tuning needs gradients, and quantized weights restrict the training path. Start experimental fine-tuning from the official FP8 or BF16 weights, get the pipeline working, and quantize for deployment afterwards.
How hy4 and A13B divide the work
This is the most asked pair inside the Hunyuan family.
| Dimension | hy4 | A13B |
|---|---|---|
| Structure | 4B dense | MoE |
| Where memory goes | The weights themselves | All total weights still land in memory |
| Inference cost | Stable and predictable | Follows active parameters, so it stays modest |
| Capability ceiling | The cost-effective tier for Chinese lightweight work | Family flagship, stronger on complex tasks |
| Official quantizations | FP8 / GPTQ-Int4 / AWQ-Int4 | FP8 / GPTQ-Int4 / official GGUF |
| Best for | Edge, embedded, high-concurrency light tasks | General chat, complex reasoning, research |
The decision rule is simple: value density per request picks the tier. If a cheaper request matters because volume is enormous, hy4's economics win. If volume is modest but each request matters a lot, A13B's capability is worth the extra footprint.
Pick by hardware
| Your hardware | Take | Why |
|---|---|---|
| 4GB of VRAM, or CPU only | Hunyuan-1.8B-Instruct | 4B at 4-bit plus KV cache gets uncomfortably tight |
| 6–8GB of VRAM | hy4 (Q4_K_M / Q5_K_M) | The comfortable zone for a 4B at 4-bit |
| 12GB and up | Higher-precision hy4 or hy3 | Depends on the value density of the task |
| 16GB and up | A13B (4-bit GGUF) | MoE total weights start around 8GB |
| Apple Silicon | hy4 MLX 4bit or A13B MLX 4bit | Unified memory makes quantization pay off directly |
| CPU-only inference | hy4 Q4/Q5 GGUF | 4B is the realistic ceiling for CPU inference |
Task mapping
| Your task | Take |
|---|---|
| Chinese customer support and Q&A | hy4 |
| Edge or embedded devices | hy4 or 1.8B |
| Intent classification, routing | Hunyuan-0.5B-Instruct |
| Code completion, small function generation | hy4 |
| Long-document analysis (hundreds of thousands of tokens) | hy4-preview (the 1M channel) |
| Multilingual translation | Hunyuan-MT-7B |
| Complex reasoning, research-style questions | A13B |
| Strict structured output | hy4 (a 4B model follows format constraints well) |
Three common mistakes
- Using hy4 where A13B belongs — pointing a 4B model at a multi-step reasoning task, then blaming "Hunyuan". That is a selection error, not a model failure.
- Treating a preview channel as an unstable one — it is a separate entry with a longer context label; what it can do depends on your workload, not on the suffix.
- Passing a HuggingFace weight name as an API model ID — local deployment uses
Hunyuan-4B-Instruct, platform calls use the real ID from the catalog, and the two naming layers never mix.
FAQ
Are hy4 and Hunyuan-4B-Instruct the same thing?
What is the difference between hy3-preview and hy3?
For hy4 at 4-bit, should I use GPTQ or AWQ?
Can a 4GB card run hy4?
Does hy4 handle images?
Can I call Hunyuan through an API?
Related reading
- hy4 model page — specification, quantization matrix and memory math
- hy4-preview channel notes — what the 1M context channel is for
- Tencent Hunyuan family overview — every Hunyuan product line in one place
- Hunyuan hy4 and A13B local deployment — deployment detail for the two main text lines
- hy4 GGUF local deployment — quantization tiers, VRAM math, and commands
- Hunyuan A13B MoE in practice — the engineering trade-offs of MoE
- hy4 vs Qwen3.8-Flash — a head-to-head on the lightweight tier
- Model directory — models callable on EXAIV today
Sources
- Tencent Hunyuan on HuggingFace
- Context and channel labels follow the live platform model directory
Download figures are HuggingFace public 30-day counters sampled in November 2026; parameters and licences follow the official model cards.