Unofficial guide
Hy4 Flash download
Last updated
Tencent publishes three Hugging Face checkpoints and points GGUF users at AngelSlim. Sizes below are sums of the files the Hugging Face tree API listed on 21 November 2026.
Instruct, FP8, and base
| Repository | Role | Tree sum |
|---|---|---|
| tencent/Hy4-flash | Instruct. 148 safetensors shards, model-00001-of-00148 through model-00148-of-00148. | 440.53 GB |
| tencent/Hy4-flash-FP8 | FP8 instruct. This is the checkpoint in the model card's vLLM and SGLang examples. | 287.09 GB |
| tencent/Hy4-flash-base | Pretrained base. The model card says the chat examples do not apply to it. | 438.84 GB |
The vLLM recipe calls the BF16 and FP8 downloads 440.5 GB and 287.1 GB, and it also describes the BF16 checkpoint as 441 GB of weights. Those match the tree sums above within the recipe's rounding. The recipe does not publish a size for the base checkpoint.
The same release is listed on ModelScope
and GitCode, which is the GitCode
URL cited with the open-weight release. The source repository is
Tencent-Hunyuan/Hy4-flash. A finetuning
note ships inside the weight repo at finetune/README.md.
Download with the Hugging Face CLI
pip install -U "huggingface_hub[cli]"
huggingface-cli download tencent/Hy4-flash-FP8
huggingface-cli download tencent/Hy4-flash
huggingface-cli download AngelSlim/Hy4-flash-GGUF The repositories were not marked gated in the Hugging Face API on 21 November 2026. The vLLM recipe still says access may require authentication, so export a token if a download or server returns 401.
Official GGUF
The model card points to AngelSlim for a GGUF conversion “for end-to-end testing.” The files in AngelSlim/Hy4-flash-GGUF on that date:
| File | Size |
|---|---|
| Hy4-flash-Q4_K_M-00001-of-00002.gguf | 97.78 GiB |
| Hy4-flash-Q4_K_M-00002-of-00002.gguf | 59.01 GiB |
| Q4_K_M pair | 156.79 GiB |
| Hy4-flash-Q3_K pair (85.69 + 59.01 GiB) | 144.70 GiB |
| Hy4-flash-IQ2_AH pair (66.72 + 59.01 GiB) | 125.73 GiB |
| Hy4-flash-SPARK-4Bit.gguf | 151.35 GiB |
| Hy4-flash-SPARK-4Bit-Fast.gguf | 148.65 GiB |
| mmproj-Hy4-flash-bf16.gguf | 0.97 GiB |
Each split quant's second shard is the same length, 63,360,616,128 bytes. AngelSlim's Mac notes
describe that ~59 GiB file as the Engram table and say it is lazily mapped. Keep both shards in
one directory. The command on the GGUF card points at the -00001-of-00002 file.
Which build fits which machine is on the GGUF page, and disk versus memory
is on requirements.
Common questions
Which Hy4 Flash checkpoint should I download?
For serving, the model card's vLLM and SGLang examples use tencent/Hy4-flash-FP8. The vLLM recipe also documents tencent/Hy4-flash, the BF16 instruct checkpoint. Download tencent/Hy4-flash-base only if you want the pretrained model. It is not instruction-tuned. For llama.cpp, use AngelSlim/Hy4-flash-GGUF.
How big is the Hy4 Flash download?
On 21 November 2026 the Hugging Face file trees summed to 440.53 GB for tencent/Hy4-flash, 287.09 GB for the FP8 instruct checkpoint, and 438.84 GB for the base checkpoint. The vLLM recipe lists the first two as 440.5 GB and 287.1 GB. Official GGUF pairs run from 125.73 GiB to 156.79 GiB.
Is Hy4 Flash gated on Hugging Face?
The Hugging Face model API reported the three Tencent checkpoints as not gated on 21 November 2026. The vLLM recipe still says a checkpoint may require a Hugging Face token, so have one ready if the server asks.
Sources
- tencent/Hy4-flash on Hugging Face Model card, architecture table, deployment commands, and license pointer.
- tencent/Hy4-flash-FP8 on Hugging Face
- tencent/Hy4-flash-base on Hugging Face Pre-trained base checkpoint, not an instruction-tuned chat model.
- AngelSlim/Hy4-flash-GGUF Official GGUF builds linked from the model card.
- Tencent-Hunyuan/Hy4-flash on GitHub Repository created 17 November 2026.
- Tencent-Hunyuan/Hy4-flash on ModelScope
- Hy4-flash on GitCode
- Tencent, “Tencent Launches Hy4 flash” Announcement dated 16 November 2026.