`hy4`

hy4 is Tencent's 4.22 billion parameter dense chat model with a native 262144 token context. Official FP8, GPTQ-Int4 and AWQ-Int4 weights ship together, with community GGUF, MLX and MNN conversions on top. This page is the local deployment reference for it.

取得 API Key 查看即時價格

hy4 (Hunyuan-4B-Instruct)

hy4 is the lightweight mainstream model on Tencent's Hunyuan text line: 4.22 billion dense parameters, a native 262144 token context, official FP8, GPTQ-Int4 and AWQ-Int4 weights published together, and community GGUF, MLX and MNN conversions on top. Its position is unambiguous — a Chinese model that runs on an ordinary machine, suited to internal networks, edge devices, and high-frequency light tasks.

The EXAIV cloud catalog is the source of truth

This page is a local deployment reference for Hunyuan hy4. Whether Hunyuan appears in the EXAIV cloud catalog depends on the live model directory. This page does not treat HuggingFace weight names as API IDs and does not provide cloud calling examples.

Specification

Item	Value
Platform model ID	hy4
HuggingFace weights	tencent/Hunyuan-4B-Instruct
Parameters	4,221,757,440 (about 4.22 billion)
Architecture	HunYuanDenseV1ForCausalLM (dense)
Layers	36
Hidden size	3072
Attention heads / KV heads	32 / 8
Intermediate size	8192
Maximum position length	262144 (256K)
Vocabulary size	120,818

Weights and quantization

Format	Provided by	Size	Fits
BF16 original	Official	about 8.4GB	Fine-tuning, FP16 inference stacks
FP8	Official	about 4.3GB	vLLM and SGLang serving
GPTQ-Int4	Official	about 2.4GB	AutoGPTQ and vLLM 4-bit routes
AWQ-Int4	Official	about 2.4GB	AutoAWQ and vLLM 4-bit routes
GGUF Q4_K_M	Community	2.61GB	The llama.cpp and Ollama default
GGUF Q5_K_M	Community	3.03GB	The quality tier when memory allows
GGUF Q8_0	Community	4.49GB	Accuracy baselines
MLX 4bit	Community	about 2.4GB	Apple Silicon
MNN	Community	varies by quantization	Mobile and embedded devices

What shipping all three quantizations means

GPTQ and AWQ are different 4-bit algorithms, and publishing both means you can test against your own inference stack instead of being locked to a single community conversion. The community GGUF line then covers every llama.cpp tier, with bartowski, mradermacher, gabriellarson, DevQuasar and Disya all publishing conversions.

Memory math

total memory  ~=  weights  +  KV cache  +  runtime overhead
KV cache      ~=  2 x 36 layers x 8 KV heads x 128 head dim x context length x 2 bytes
              ~=  144KB x context length

Combination	Estimated footprint
Q4_K_M with a 4K context	about 3.7GB
Q4_K_M with a 32K context	about 7.7GB
Q5_K_M with an 8K context	about 4.6GB
FP8 with a 4K context	about 5.4GB

Deployment paths

Situation	Path
Single consumer GPU	Official GPTQ-Int4 or AWQ-Int4 with vLLM
CPU, Mac, or no GPU	Community GGUF or MLX with llama.cpp or Ollama
Apple Silicon	The mlx-community 4bit build with mlx-lm
Edge and embedded	The taobao-mnn MNN release
Private inference service	Official FP8, keeping quality headroom

Where it fits

Everyday Chinese conversations, support and summarising
High-frequency light tasks: rewriting, routing, classification
Data processing inside internal networks and offline environments
The front half of a RAG pipeline: rewriting, routing, snippet summarising
Edge and embedded devices

Where it does not fit

Multi-hop reasoning and hard mathematics
Long-form synthesis and report writing
Tasks that need frontier coding ability
High-volume high-quality content creation

For those, look at Hunyuan A13B (the MoE flagship) or a cloud model.

hy4 GGUF local deployment — quantization tiers, memory math and full commands
hy4 model selection guide — how it divides work with hy3 and A13B
Tencent Hunyuan family overview — every Hunyuan product line
Running hy4 with Ollama — wiring it into Dify, Cursor and other tools
hy4 local RAG in practice — pointing a knowledge base at it
VRAM calculation guide — the general memory math

模型規格

API 模型名	`hy4`
廠商	腾讯混元
上下文視窗	—
類型	chat
開源狀態	Released under the vendor licence. Official weights cover BF16, FP8, GPTQ-Int4 and AWQ-Int4; community conversions cover GGUF, MLX and MNN.

價格按 Token 點數即時計費，即時單價見模型價格頁。

模型亮點

4.22 billion dense parameters, so every token is fully computed and behaviour is predictable
Native 262144 (256K) context, the longest window on Hunyuan's lightweight line
Official FP8, GPTQ-Int4 and AWQ-Int4 weights mean 4-bit does not depend on community work
Multiple community GGUF publishers, with Q4_K_M at 2.61GB so a 6GB card runs it

典型使用場景

1Chinese-first assistants

Everyday Chinese question answering, support and summarising, which is where a 4B-class model hits its value sweet spot.

2On-premise and edge deployment

Private deployments where data cannot leave the environment, or edge devices with limited memory.

3The front half of a retrieval pipeline

Query rewriting, intent routing and snippet summarising, working alongside vector search.

快速開始

本模型透過全站統一的 OpenAI 相容 API 提供（Bearer 驗證）。端點： POST /api/ai/v1/chat/completions · API 文件 →

Python（OpenAI SDK）

python

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url=" echo exaiv.com;; *) echo https://exaiv.com;; esac)/api/ai/v1"
)

response = client.chat.completions.create(
    model="hy4",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "你好，介绍一下你自己"}
    ]
)
print(response.choices[0].message.content)

串流輸出

python

stream = client.chat.completions.create(
    model="hy4",
    messages=[{"role": "user", "content": "写一首关于秋天的小诗"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

cURL

bash

curl  echo exaiv.com;; *) echo https://exaiv.com;; esac)/api/ai/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy4",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": false
  }'

示例中的 API 域名已按当前站点动态填入（构建时注入的网关地址），无需手动替换，可直接使用。

常見問題

+hy4 (Hunyuan-4B-Instruct) 的上下文視窗是多大？

音訊等特殊模型，不適用文字上下文視窗。

+怎麼計費？

按 Token 用量以點數計費，快取命中更優。即時單價見模型價格頁。

+需要註冊原廠商帳號嗎？

不需要。一組 AI Gateway API Key 即可呼叫本模型及站內全部其他模型。

+權重開源嗎？

Released under the vendor licence. Official weights cover BF16, FP8, GPTQ-Int4 and AWQ-Int4; community conversions cover GGUF, MLX and MNN.

立即使用 hy4

一組 API Key，全部模型。點數永不過期。

免費註冊

hy4 (Hunyuan-4B-Instruct)

The EXAIV cloud catalog is the source of truth

Specification

Item	Value
Platform model ID	hy4
HuggingFace weights	tencent/Hunyuan-4B-Instruct
Parameters	4,221,757,440 (about 4.22 billion)
Architecture	HunYuanDenseV1ForCausalLM (dense)
Layers	36
Hidden size	3072
Attention heads / KV heads	32 / 8
Intermediate size	8192
Maximum position length	262144 (256K)
Vocabulary size	120,818

Weights and quantization

Format	Provided by	Size	Fits
BF16 original	Official	about 8.4GB	Fine-tuning, FP16 inference stacks
FP8	Official	about 4.3GB	vLLM and SGLang serving
GPTQ-Int4	Official	about 2.4GB	AutoGPTQ and vLLM 4-bit routes
AWQ-Int4	Official	about 2.4GB	AutoAWQ and vLLM 4-bit routes
GGUF Q4_K_M	Community	2.61GB	The llama.cpp and Ollama default
GGUF Q5_K_M	Community	3.03GB	The quality tier when memory allows
GGUF Q8_0	Community	4.49GB	Accuracy baselines
MLX 4bit	Community	about 2.4GB	Apple Silicon
MNN	Community	varies by quantization	Mobile and embedded devices

What shipping all three quantizations means

Memory math

total memory  ~=  weights  +  KV cache  +  runtime overhead
KV cache      ~=  2 x 36 layers x 8 KV heads x 128 head dim x context length x 2 bytes
              ~=  144KB x context length

Combination	Estimated footprint
Q4_K_M with a 4K context	about 3.7GB
Q4_K_M with a 32K context	about 7.7GB
Q5_K_M with an 8K context	about 4.6GB
FP8 with a 4K context	about 5.4GB

Deployment paths

Situation	Path
Single consumer GPU	Official GPTQ-Int4 or AWQ-Int4 with vLLM
CPU, Mac, or no GPU	Community GGUF or MLX with llama.cpp or Ollama
Apple Silicon	The mlx-community 4bit build with mlx-lm
Edge and embedded	The taobao-mnn MNN release
Private inference service	Official FP8, keeping quality headroom

Where it fits

Everyday Chinese conversations, support and summarising
High-frequency light tasks: rewriting, routing, classification
Data processing inside internal networks and offline environments
The front half of a RAG pipeline: rewriting, routing, snippet summarising
Edge and embedded devices

Where it does not fit

Multi-hop reasoning and hard mathematics
Long-form synthesis and report writing
Tasks that need frontier coding ability
High-volume high-quality content creation

For those, look at Hunyuan A13B (the MoE flagship) or a cloud model.

hy4 GGUF local deployment — quantization tiers, memory math and full commands
hy4 model selection guide — how it divides work with hy3 and A13B
Tencent Hunyuan family overview — every Hunyuan product line
Running hy4 with Ollama — wiring it into Dify, Cursor and other tools
hy4 local RAG in practice — pointing a knowledge base at it
VRAM calculation guide — the general memory math

模型規格

API 模型名	`hy4`
廠商	腾讯混元
上下文視窗	—
類型	chat
開源狀態	Released under the vendor licence. Official weights cover BF16, FP8, GPTQ-Int4 and AWQ-Int4; community conversions cover GGUF, MLX and MNN.

價格按 Token 點數即時計費，即時單價見模型價格頁。

模型亮點

4.22 billion dense parameters, so every token is fully computed and behaviour is predictable
Native 262144 (256K) context, the longest window on Hunyuan's lightweight line
Official FP8, GPTQ-Int4 and AWQ-Int4 weights mean 4-bit does not depend on community work
Multiple community GGUF publishers, with Q4_K_M at 2.61GB so a 6GB card runs it

典型使用場景

1Chinese-first assistants

Everyday Chinese question answering, support and summarising, which is where a 4B-class model hits its value sweet spot.

2On-premise and edge deployment

Private deployments where data cannot leave the environment, or edge devices with limited memory.

3The front half of a retrieval pipeline

Query rewriting, intent routing and snippet summarising, working alongside vector search.

快速開始

本模型透過全站統一的 OpenAI 相容 API 提供（Bearer 驗證）。端點： POST /api/ai/v1/chat/completions · API 文件 →

Python（OpenAI SDK）

python

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url=" echo exaiv.com;; *) echo https://exaiv.com;; esac)/api/ai/v1"
)

response = client.chat.completions.create(
    model="hy4",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "你好，介绍一下你自己"}
    ]
)
print(response.choices[0].message.content)

串流輸出

python

stream = client.chat.completions.create(
    model="hy4",
    messages=[{"role": "user", "content": "写一首关于秋天的小诗"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

cURL

bash

curl  echo exaiv.com;; *) echo https://exaiv.com;; esac)/api/ai/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy4",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": false
  }'

示例中的 API 域名已按当前站点动态填入（构建时注入的网关地址），无需手动替换，可直接使用。

常見問題

+hy4 (Hunyuan-4B-Instruct) 的上下文視窗是多大？

音訊等特殊模型，不適用文字上下文視窗。

+怎麼計費？

按 Token 用量以點數計費，快取命中更優。即時單價見模型價格頁。

+需要註冊原廠商帳號嗎？

不需要。一組 AI Gateway API Key 即可呼叫本模型及站內全部其他模型。

+權重開源嗎？

Released under the vendor licence. Official weights cover BF16, FP8, GPTQ-Int4 and AWQ-Int4; community conversions cover GGUF, MLX and MNN.

立即使用 hy4

一組 API Key，全部模型。點數永不過期。

免費註冊