Skip to content
TrustList
All launches
AI model launch

Hy4 Flash

Tencent's open-weight Hy4 Flash, listed 16 November 2026: 220B-parameter mixture of experts with 12B active, 1,048,576-token context, text and image in; regular price $0.07 input and $0.281 output per million tokens.

Launched 2026-11-16AI model

About this launch

Tencent describes Hy4 Flash as the most cost-effective model in its Hy family so far. It was listed on OpenRouter on 16 November 2026, and the open weights appeared on Hugging Face the next day. It reads text and images and answers in text.

## Size and design

The model is a mixture of experts with 220 billion parameters in total, of which about 12 billion are active for each token (11.3 billion for text plus 0.43 billion for vision). The total counts 155.5 billion main parameters and 63.5 billion in an n-gram embedding table. Each token uses 16 of 512 routed experts plus one shared expert. Attention mixes sparse and sliding-window layers, and the context window is 1,048,576 tokens. A small 0.85-billion-parameter drafter is built in to speed up decoding on long contexts.

## What Tencent says it is for

The model card recommends it for everyday productivity work such as deep search, document processing, slide creation, simple front-end demos and computer use, and points to the larger Hy4 for long-horizon software engineering. Tencent reports a score of 42.1 on its own WorkIndex tasks against 39.7 for Hy4 Preview, at about one thirteenth of the cost per task, and a user-satisfaction score of 2.79 out of 4 in an internal blind test with 85 of its own staff. These are Tencent's own figures.

## Weights, licence and price

Weights for the instruct model, an FP8 version and a base model are published on Hugging Face, ModelScope and GitCode, with community GGUF files and setup guides for vLLM and SGLang. The licence is the Tencent Hy License 1.0, which is derived from Apache 2.0 but is not Apache: a business that gives third parties API access to the model needs a separate written licence if its worldwide revenue is above US$50 million a year, or if the service has more than 100 million monthly users. Internal use is excluded.

On OpenRouter, Tencent is the only host. The regular rates are $0.07 per million input tokens and $0.281 per million output tokens; OpenRouter currently shows a 50% launch discount, which is not the price recorded here.

## Key facts

**Maker:** Tencent **Listed:** 16 November 2026 (OpenRouter); open weights on Hugging Face 17 November 2026 **Inputs:** text, images **Context window:** 1,048,576 tokens **Parameters:** 220B total, 12B active per token **Price per million tokens:** $0.07 input, $0.281 output (regular; 50% launch discount not recorded) **Access:** OpenRouter; open weights on Hugging Face, ModelScope and GitCode **Licence:** Tencent Hy License 1.0 (custom)

Not yet independently verified: Only Tencent's own Hugging Face model card and licence file and one tracker, OpenRouter, back these facts; Tencent's own web page for the model did not load for us. The 17 November date for the open weights comes from the Hugging Face repository history, not from a statement by Tencent. The task, cost and user-satisfaction figures on the card are Tencent's own and have not been checked independently, so no scores are stored here. The regular price is recorded, not the launch discount. We will update this when it can be confirmed, and remove this note.

Discussion

Questions & answers about Hy4 Flash.

No questions yet. Be the first to ask about Hy4 Flash.