[Models](/models)/tencent/Hy4-flash

Tencent Hy License 1.0 (Apache 2.0 based; a separate license is needed for large model-as-a-service businesses)

# Hy4-flash

by tencent

Tencent's open-weight image and text model: 220B mixture-of-experts with about 12B active per token, a SigLIP2 vision encoder and a 1M-token context

386likes

220B total (155.5B backbone + 63.5B Engram + 0.43B vision encoder + 0.85B drafter); about 12B active per tokenparams

[HuggingFace](https://huggingface.co/tencent/Hy4-flash)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`tencent/Hy4-flash`

## Deploy Hy4-flash

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

Hy4 flash is a vision-language model from Tencent, released on Hugging Face on November 17, 2026 under the Tencent Hy License. It takes text and images as input and writes text, with a context window of 1,048,576 tokens. Tencent calls it the most cost-effective Hy model so far: 220B parameters in total, counting 63.5B of learned n-gram (Engram) tables, with about 11.3B text parameters and the 0.43B vision encoder active for each token.

The vision side is a SigLIP2-So400m encoder adapted at 512-pixel patches and joined to the language backbone by an MLP projector. The backbone has 42 layers, 41 of them mixture-of-experts, where each token uses 16 of 512 routed experts plus one shared expert, and attention interleaves sparse and sliding-window layers. A built-in 0.85B drafter, DFly2, speeds up decoding by speculation.

On Tencent's internal WorkIndex, which includes audio and video processing and visual understanding tasks, it scores 42.1 against 39.7 for Hy4 preview at about 1/13 of the cost per task. For a media pipeline, it reads frames, slides and scanned pages and writes descriptions or structured fields you can index and search.

## Architecture

Mixture-of-experts language backbone (42 layers: 1 dense, 41 MoE; 16 of 512 routed experts plus 1 shared expert per token; hidden size 3,072; multi-head latent attention with sparse and sliding-window layers; interleaved MRoPE for image positions) with a vision encoder adapted from SigLIP2-So400m-Patch16-Res512 through an MLP projector. Learned bigram and trigram (Engram) tables add 63.5B parameters, and a 5-layer, 0.85B DFly2 drafter supports speculative decoding.

## Mixpeek SDK Integration

IngestRetrieve

\# Hy4 flash is not a managed extractor. Serve it yourself (Tencent's vLLM
# recipe runs the FP8 weights on 8 GPUs with --tensor-parallel-size 8), caption
# each frame, and store the caption as searchable text in a Mixpeek collection.
import base64, requests
from openai import OpenAI

llm = OpenAI(base\_url="http://localhost:8000/v1", api\_key="unused")

def caption(path):
    b64 = base64.b64encode(open(path, "rb").read()).decode()
    r = llm.chat.completions.create(
        model="hy4-flash",
        messages=\[{"role": "user", "content": \[
            {"type": "image\_url", "image\_url": {"url": f"data:image/jpeg;base64,{b64}"}},
            {"type": "text", "text": "Describe what is happening in this frame in one sentence."},
        \]}\],
        max\_tokens=120,
    )
    return r.choices\[0\].message.content

requests.post(
    "https://api.mixpeek.com/v1/namespaces/ns\_your\_namespace/documents/upsert",
    headers={"Authorization": "Bearer API\_KEY"},
    json={"collection\_id": "col\_frame\_captions",
          "documents": \[{"document\_id": "clip-042-t0012",
                         "payload": {"caption": caption("clip-042\_0012.jpg"),
                                     "video": "clip-042.mp4", "start\_time": 12.0}}\]},
)

## Capabilities

*   Image and text input, text output, with a 1M-token context window
*   About 12B active parameters per token out of 220B in total
*   Built-in DFly2 drafter for speculative decoding; Tencent reports over 150 tokens/s with 4-bit weights
*   Released as instruct, FP8 instruct and pre-trained base weights
*   Serves through vLLM or SGLang with an OpenAI-compatible API

## Use Cases on Mixpeek

Captioning sampled video frames so footage can be searched by what is on screen

Reading slides, screenshots and scanned pages into text for document search

Extracting structured fields from product or ad images for filtering

Answering questions over retrieved images in a multimodal RAG pipeline

## Benchmarks

Dataset

Metric

Score

Source

WorkIndex (Tencent internal, 8 task sets)

Score

42.1

Model card: tencent/Hy4-flash (self-reported; Hy4 preview 39.7, Hy4 51.4)

Blind user evaluation (85 users, 259 tasks)

Satisfaction (of 4)

2.79

Model card (self-reported; Claude Haiku 5.5: 2.81)

## Performance

Input SizeImages and text, up to 1,048,576 tokens of context

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

Tencent's serving recipes run the FP8 weights with tensor parallelism across 8 GPUs. Tencent reports decoding above 150 tokens/s with 4-bit weights and the DFly2 drafter. We have not measured it.

## Common Pipeline Companions

[

Qwen/Qwen3.8-27B

A smaller open vision-language model that fits on far less hardware

](/model/Qwen/Qwen3.8-27B)

## Frequently Asked Questions

### What is Hy4 flash?

An open-weight vision-language model from Tencent, released November 17, 2026. It reads text and images and writes text, has a 1M-token context, and is a 220B mixture-of-experts model with about 12B parameters active per token.

### Can Hy4 flash understand images and video?

It takes images natively through a SigLIP2-based vision encoder. To work with video, sample frames and pass them as images; Tencent's internal WorkIndex includes audio and video processing tasks, but the card does not document direct video input.

### What hardware does Hy4 flash need?

Tencent's vLLM and SGLang recipes serve the FP8 weights with tensor parallelism across 8 GPUs. Only about 12B parameters are active per token, so it decodes fast once loaded, but all 220B parameters still have to fit in GPU memory.

### Can I use Hy4 flash commercially?

Yes, under the Tencent Hy License 1.0, which is based on Apache 2.0. A company that offers general model access to third parties as a model-as-a-service business needs a separate written license from Tencent above the thresholds in section 10, such as US$50 million in yearly revenue or 100 million monthly active users. Read the license for the exact terms.

### How do I use Hy4 flash with Mixpeek?

Serve it with vLLM or SGLang, send it sampled video frames or images to caption, and upsert the captions into a Mixpeek collection. A retriever then searches those captions by meaning, alongside any image embeddings in the same namespace.

## Explore on Mixpeek

[

Browse all models

Search the full model catalog

](/models)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

Organizationtencent

Retriever\-

Parameters220B total (155.5B backbone + 63.5B Engram + 0.43B vision encoder + 0.85B drafter); about 12B active per token

LicenseTencent Hy License 1.0 (Apache 2.0 based; a separate license is needed for large model-as-a-service businesses)

Downloads/moN/A

Likes386

[

### Research Paper

Hy4 flash model card (Tencent, November 2026)

arxiv.org](https://huggingface.co/tencent/Hy4-flash)

### Build a pipeline with Hy4-flash

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)