Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
Tencent Hy4 Preview Is Real Weights. Not a GX10 Sit.
tencent/Hy4-preview lastModified 2026-08-28T14:58 UTC. License apache-2.0. 131 safetensor shards. Hugging Face usedStorage 1,560,008,337,568 bytes.
Key takeaways
- tencent/Hy4-preview lastModified 2026-08-28T14:58 UTC. License apache-2.0. 131 safetensor shards. Hugging Face usedStorage 1,560,008,337,568 bytes.
- Card: 770B total / 49B active. 78 layers, 256 routed experts plus 1 shared, top-8. Native MTP 10B / 0.7B active. Context 1M. model_type hy_v4.
- Official FP8 tencent/Hy4-preview-FP8 usedStorage 813,790,680,536 bytes. No public GGUF: tencent/Hy4-preview-GGUF and unsloth/Hy4-preview-GGUF both 401. The card names vLLM and SGLang. I did not load this. Not a sit.
Local LLMs on NVIDIA Spark / ASUS GX10
Tencent posted Hy4-preview last Friday. The instrument is a Hugging Face card, last updated Aug 28 at 14:58 UTC. tencent/Hy4-preview. License: Apache 2.0. 131 safetensor shards. I pulled the API: usedStorage is 1,560,008,337,568 bytes.
The card's own count: 770B total parameters, 49B activated per token. 78 layers. First layer is a dense FFN; the other 77 are MoE with 256 routed experts and 1 shared expert, top-8. Native MTP: 10B total, 0.7B activated. Context 1M. model_type is hy_v4. Architecture HYV4ForCausalLM.
I am not quoting their benchmark appendix. I am not inventing tokens per second.
The fit test is still the file size. Official BF16-class weights are about 1.56 TB. Official FP8 (tencent/Hy4-preview-FP8, lastModified 2026-08-28T14:58 UTC) usedStorage is 813,790,680,536 bytes. Neither of those fits 128GB unified.
There is no public GGUF. I hit tencent/Hy4-preview-GGUF and unsloth/Hy4-preview-GGUF. Both 401. The card names vLLM and SGLang for deploy. I am not standing those up.
I run Grok when it earns it, and a GX10 when I want the weights in the room. This family is on the far side of that split. If a later quant actually sits in 128GB and I load it, that note comes next. Not this card, and not a download I did not start.
Frequently asked questions
Yes. The Hugging Face card tencent/Hy4-preview was last updated Aug 28, 2026 at 14:58 UTC. License is Apache 2.0. I am writing it now that wikiwayne.com HTTPS is back.
No. Official usedStorage is 1,560,008,337,568 bytes. Official FP8 is 813,790,680,536. There is no public GGUF to even argue about. File size is not a sit. I did not load this on the GX10.
The card is tencent/Hy4-preview, model_type hy_v4. Tencent's own table: 770B / 49B active. I am not rewriting an older Hunyuan sit, and I am not quoting their benchmark appendix.
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.