Unofficial guide
Hy4 Flash
Last updated
Hy4 Flash is Tencent's open-weight multimodal mixture-of-experts model for everyday work. The weights, a vLLM recipe, an SGLang command, and a GGUF repo are all public. This page keeps the figures that those sources actually print.
- Total parameters
- 220B on the model card
- Active per token
- 12B, text 11.3B
- Context
- 1,048,576 tokens
- Input / output
- Text and image in, text out
- Weights
- Open, Tencent Hy License 1.0
- First published
- 16–17 November 2026
What it is
Tencent's model card calls Hy4 Flash its most cost-effective Hy model so far: 220 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters per token. It accepts text and images and generates text. The same card's table is tighter than the headline. The language backbone, excluding Engram, the vision encoder, and the DFly2 drafter, is 155.5 billion parameters. Engram is 63.5 billion. The vision encoder is 0.43 billion. DFly2 is 0.85 billion. Those lines sum to 220.28 billion. Active parameters are 11.3 billion for text, plus the 0.43 billion vision encoder when an image is in the prompt.
Hugging Face's model API reports 220,251,201,958 parameters for tencent/Hy4-flash.
Tencent's announcement dated 16 November 2026 rounds the same release to 224 billion total
parameters, including 64 billion Engram parameters, and 12 billion active. The repository on
Hugging Face was created on 17 November 2026, and the GitHub repository the same day.
The backbone has 42 layers: one dense feed-forward layer and 41 mixture-of-experts layers, with identity Hyper-Connections across four residual streams. Hidden size is 3,072 and the vocabulary is 120,832. Each expert layer has 512 routed experts and one shared expert. A token activates 16 routed experts. Those experts run in a latent dimension of 1,536, which the card calls LatentMoE. Attention alternates sparse IndexPool layers and sliding-window layers. The pattern is two sparse layers, then ten repeats of three sliding-window layers and one sparse layer: 12 sparse layers and 30 sliding-window layers. Sparse layers can select up to 2,048 tokens. The sliding window is 512 tokens. Engram adds learned bigram and trigram embeddings, looked up by hash, on layers 1, 11, 21, and 31. The vision tower is adapted from SigLIP2-So400m-Patch16-Res512 and joined through an MLP. DFly2 is a 5-layer drafter that reuses the target model's KV cache and sparse indices.
Tencent says the KV cache is 1/24 the size of Hy3 and 1/6 the size of Hy4, and that their serving stack (megakernels, 4-bit quantization, and DFly2) decodes past 150 tokens per second. That speed is their claim for that stack, not a measurement from this site. They recommend Hy4 Flash for deep search, document processing, slides, simple frontend demos, and computer use, and the full Hy4 model for long-horizon software engineering.
Three ways to use it
hy4-flash. List price is on their pricing page. Serve it Official vLLM and SGLang Docker images. The published examples use eight GPUs. GGUF AngelSlim's quants. Published speeds are for 128 GB machines, and llama.cpp support was still a pull request on 21 November 2026. Quick start
-
If you want answers without a download, use the hosted API. The model id on
TokenHub is
hy4-flash. -
If you have a multi-GPU server, follow the vLLM or SGLang
command from the model card. Both sample commands use
tencent/Hy4-flash-FP8and--tensor-parallel-size 8or--tp 8. - If you are on a 128 GB workstation, read the GGUF page before downloading. Q4_K_M is the build AngelSlim calls the safe default, and the file pair is 156.79 GiB.
The model card's recommended sampling settings are temperature=0.9 and
top_p=1.0. On the instruct checkpoints, reasoning_effort is
"high" by default and "no_think" for a direct answer. The base
checkpoint, tencent/Hy4-flash-base, is a pretrained model, not a chat model.
Where the weights are
| Checkpoint | What it is | Listed size |
|---|---|---|
| tencent/Hy4-flash | Instruct, BF16 weights | 440.53 GB |
| tencent/Hy4-flash-FP8 | FP8 instruct | 287.09 GB |
| tencent/Hy4-flash-base | Pretrained base, not a chat model | 438.84 GB |
| AngelSlim/Hy4-flash-GGUF | Official GGUF quants | 125.73–156.79 GiB |
The same card also links ModelScope and GitCode. Mirrors checked for this guide: ModelScope and GitCode. File-by-file sizes and the GGUF shard names are on the download page.
Common questions
What is Hy4 Flash?
Hy4 Flash, also written Hy4 flash and Hy4-flash, is an open-weight multimodal mixture-of-experts model from Tencent's Hy team. The model card says it takes text and image input, generates text, and has a context window of 1,048,576 tokens.
How many parameters does Hy4 Flash have?
The model card says 220 billion total parameters, including about 64 billion Engram parameters, with 12 billion active per token. Its table breaks that into 155.5 billion backbone parameters excluding Engram, the vision encoder, and DFly2, plus 63.5 billion Engram, 0.43 billion vision, and 0.85 billion DFly2. Text activation is 11.3 billion, plus the 0.43 billion vision encoder when an image is used. Hugging Face reports 220,251,201,958 parameters. Tencent's 16 November 2026 announcement rounds the total to 224 billion.
Can I run Hy4 Flash on a laptop?
The BF16 checkpoint is about 440.5 GB and the vLLM recipe targets eight Hopper or Blackwell data-center GPUs. AngelSlim's published GGUF speeds are for a 128 GB Mac M5 Max and a 128 GB RTX Spark. This guide has no sourced setup for a 16 GB or 24 GB laptop.
What license is Hy4 Flash under?
Tencent Hy License, Version 1.0. It is derived from the Apache License 2.0 and adds a separate-license requirement for very large model-as-a-service operators. Read the license page for the thresholds.
Sources
- tencent/Hy4-flash on Hugging Face Model card, architecture table, deployment commands, and license pointer.
- Tencent, “Tencent Launches Hy4 flash” Announcement dated 16 November 2026.
- Hugging Face model API for tencent/Hy4-flash Reports 220,251,201,958 parameters. Repository created 17 November 2026.
- Tencent-Hunyuan/Hy4-flash on GitHub Repository created 17 November 2026.
- Tencent Hy research page for Hy4 flash Official research URL. The page is a web app; specs below come from the model card and GitHub README.
- vLLM recipe for tencent/Hy4-flash
- AngelSlim/Hy4-flash-GGUF Official GGUF builds linked from the model card.
- Tencent Hy License, Version 1.0 Apache License 2.0 terms plus Section 10 for large MaaS operators.
- Tencent Cloud, Hy API Guide TokenHub model ids, context, and a chat-completions example. Page updated 16 November 2026.
- Tencent Cloud, TokenHub model pricing English list prices. Page updated 19 November 2026.