INDEPENDENT TECHNOLOGY JOURNALISMENGLISH / हिंदी
AI & Tools

Tencent's Hy4-flash is trending on Hugging Face: a 220B-parameter model that answers questions about images

Tencent's Hy4-flash, a multimodal mixture-of-experts model that reads images and answers questions in text, is drawing attention in Hugging Face's trending section. It pairs 220B total parameters with just 12B active per token.

Editorial illustration for: Tencent Hy4-flash: images in, answers out
LUCIDWIRE editorial illustration

A Tencent model called Hy4-flash is trending on Hugging Face, and its model card explains why developers are looking: it accepts both text and image inputs, generates text answers, and supports a context window of up to 1 million tokens, while keeping costs low.

In practical terms, this is a vision-language model. You can show it a picture and ask questions about it, and it replies in text. Tencent describes Hy4-flash as its most cost-effective Hy model so far. It has 220B total parameters, including 64B "Engram" parameters, but activates only around 12B parameters per token. That mixture-of-experts design is what makes a very large model cheaper to run, since each request uses only a small slice of the whole network.

Under the hood, the 42-layer language backbone combines one dense layer with 41 mixture-of-experts layers, and a vision encoder adapted from SigLIP2 feeds images into it. Three changes set it apart from the earlier Hy4: hybrid attention that mixes sparse and sliding-window attention, a LatentMoE setup where routed experts work in a smaller latent space, and learned bigram and trigram embeddings the company calls Engram. A built-in drafter module, DFly2, supports speculative decoding to speed up long responses.

Two cautions are worth keeping. The benchmark numbers on the card come from Tencent's own evaluations, and the card was updated on 19 November, so this is a recently revised listing rather than a brand-new release. Still, for developers hunting for a cheaper multimodal model with a long context window, Hy4-flash is now one of the more visible options on Hugging Face.

Source: Hugging Face ↗

Updated: 20/11/2026, 11:33:21 am IST

Related stories