Skip to main content
Watch: I Built an AI App Without Leaving Databricks

Tencent - Mixture of experts - verified

Run Hy4 preview locally

Hy4 preview is a 780B mixture-of-experts (49B active) model from Tencent with a 1.05M-token context window, released under Apache 2.0 (commercial use is allowed); the smallest listed download is 433.4 GB, the 4bit MLX build from mlx-community/Hy4-preview-4bit.

Official weights: tencent/Hy4-preview

Total params
780B
Active params
49B
Context
1.05M
Smallest download
433.4 GB

Fits in

Smallest download at most 85% of memory

None of the tiers here. The smallest listed download is 433.4 GB, more than 85% of 256 GB.

Downloads

The official weights and every quantization listed for this model, with file sizes from Hugging Face.

Hy4 preview downloads and file sizes
DownloadFormatSizeRepo
Official weightsBF161,560.0 GBtencent/Hy4-preview
Q4_K_MGGUF467.3 GBAngelSlim/Hy4-preview-GGUF
4bitSmallest listedMLX433.4 GBmlx-community/Hy4-preview-4bit

GB is decimal: 1 GB = 10^9 bytes. Add memory for the context and the operating system on top of the file size.

Run it

Copied verbatim from the page linked under each command. Server commands from model cards often assume a multi-GPU node; check the flags before you run them.

vLLM

Terminal
docker run --gpus all \
  -p 8000:8000 \
  --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 \
    --tensor-parallel-size 8 \
    --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' \
    --attention-backend FLASHMLA_SPARSE \
    --tool-call-parser hy_v4 \
    --reasoning-parser hy_v4 \
    --enable-auto-tool-choice \
    --port 8000 \
    --served-model-name hy4-preview

Source: huggingface.co/tencent/Hy4-preview

SGLang

Terminal
docker pull lmsysorg/sglang:hy4-preview

docker run --gpus all --ipc=host -p 8000:8000 lmsysorg/sglang:hy4-preview \
  python3 -m sglang.launch_server \
    --model tencent/Hy4-preview-FP8 \
    --tp-size 8 \
    --reasoning-parser auto \
    --tool-call-parser auto \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --port 8000 \
    --served-model-name hy4-preview

Source: huggingface.co/tencent/Hy4-preview

Not sure which runtime to use? Compare local LLM runtimes by platform, API and features.

License

License
Apache 2.0apache-2.0
Commercial use
Commercial use: YesCommercial use is allowed.
Gated download
No. The files download directly.

Worth knowing

The card describes this as an early preview that is known to over-reason, and its official serving commands use the companion tencent/Hy4-preview-FP8 repo with dedicated hy4-preview container images.

Released (Hugging Face repo creation date).

API pricing and specsHosted prices per 1M tokens for Hy4 preview, if you would rather not run it yourself.

FAQ

How big is the Hy4 preview download?

The official BF16 weights in tencent/Hy4-preview are 1,560.0 GB. The smallest listed download is 433.4 GB, the 4bit MLX build from mlx-community/Hy4-preview-4bit. Plan for more memory than the file size: the context (KV cache) and the operating system need room too.

Can I use Hy4 preview commercially?

Hy4 preview is released under Apache 2.0, and commercial use is allowed. The license text is at https://www.apache.org/licenses/LICENSE-2.0.

How much memory do I need for Hy4 preview?

Its smallest listed download is 433.4 GB, more than 85% of 256 GB, so it needs a multi-GPU server or a machine with more memory than the largest tier here.

What is the context length of Hy4 preview?

1,048,576 tokens. A longer context needs more memory for the KV cache, so you can run with a shorter one when memory is tight.

Sources

Last verified from the pages below.

Newsletter

New open weights, every week

One email per week. New model releases, open-weight drops and what they mean for what you build.

Keep exploring

Best local AI models

Our curated picks by hardware tier, with benchmarks and the reasoning.

Model API pricing and specs

First-party prices per 1M tokens when you would rather call an API.

AI benchmark leaderboards

Sourced scores, for when you need to compare quality.

Setup guides

Point your coding tools at the model you picked.