Tencent launches Hy4 Flash, a 224B cost-tier model at about $0.07 per million input tokens
Tencent launched Hy4 Flash on Monday, November 16, a smaller, cheaper sibling of its Hy4 flagship — a 224-billion-parameter model priced at about $0.07 per million input tokens.
📊 Scorecard: our Bullish call on $0700.HK is scored from HK$390.40, the last close before it ran. Its first result comes after the next close.

The News
Tencent launched Hy4 Flash on Monday, November 16, a smaller and cheaper sibling of its Hy4 flagship model, which was released on November 4. Hy4 Flash is a 224-billion-parameter mixture-of-experts model, with 12 billion parameters active for each token it processes. It has a 1-million-token context window and accepts multimodal input, meaning it can take in text and other kinds of input such as images. The model is live in Tencent's WorkBuddy and Yuanbao products and through Tencent Cloud TokenHub.
The headline is the price. Tencent's published API list prices are about $0.07 per million input tokens and $0.281 per million output tokens. Input tokens are the words and data you feed the model; output tokens are the words it writes back, and they usually cost more. At these prices, Hy4 Flash is aimed squarely at developers and businesses that want a capable model without paying flagship rates for routine work.
Tencent says that on its internal WorkIndex benchmark, Hy4 Flash cost one-twentieth as much per task as the full Hy4 model — in other words, about 95% less. That is a vendor-reported internal figure, not an independent evaluation, so it should be read as Tencent's own claim rather than a verified fact. Tencent also says the model's weights — the files that let others run it themselves — and an OpenRouter listing are promised, though they have not yet arrived. Tencent's shares closed at HK$390.40 on Monday, November 16, up about 1%.
Hy4 Flash is priced at about $0.07 per million input tokens — a cost-tier model aimed at routine work.
Why It Matters
The price of running AI models is becoming one of the most important numbers in the industry, because it decides what gets built. A flagship model may be brilliant, but if it costs too much per request, developers will only use it for special jobs. A cheaper model that is good enough for routine work opens the door to far more uses — customer service, document handling, everyday coding help — and that is where the volume is. Hy4 Flash is Tencent's bet that the cost tier is where the next wave of users lives.
The design of the model fits that bet. With 224 billion parameters in total but only 12 billion active for each token, Hy4 Flash uses the mixture-of-experts trick: keep a large pool of specialist knowledge, but switch on only the experts needed for each word. That keeps running costs down while keeping the model capable. A 1-million-token context window adds to the appeal, letting the model swallow very long documents in one go instead of working in small chunks.
The timing matters too. Hy4 Flash arrives less than two weeks after the Hy4 flagship, which shows Tencent is moving fast to cover both ends of the market: the flagship for the hardest jobs, the cheaper sibling for everything else. The model is already live in WorkBuddy and Yuanbao, so it is not a paper launch. If the promised weights and OpenRouter listing follow, outside developers will be able to test the cost claims for themselves, which is the moment this story gets its real verdict.
Analyst's Take
Our call is Bullish. A capable model at about $0.07 per million input tokens is a genuine step in the race to make AI cheap enough for everyday business use, and Tencent is shipping it inside products people already use, not just announcing it. If the price holds up in real use, Hy4 Flash gives Tencent a sharp tool for winning developers who watch every cent.
The honest caveat is that the most eye-catching number — one-twentieth the cost per task of the full Hy4 — comes from Tencent's own internal benchmark, not from an independent test. Vendor benchmarks are designed to show a product in its best light. The weights and the OpenRouter listing are also only promised so far, so outside developers cannot yet check the model for themselves. Until independent results arrive, the cost claim deserves a grain of salt.
Even with that caution, the direction is right. Cheaper models expand the market: tasks that were too expensive to automate become worth doing, and more developers can afford to build with AI. Tencent is following a proven playbook — ship the flagship first, then the affordable sibling that reaches far more users. Reader poll: Bullish or bearish on 0700.HK? Vote: Bullish, Bearish, Mixed, Wait for proof.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is a mixture-of-experts model?
It is a model built from many specialist sub-models, or experts, where only a few are switched on for each word or task, which keeps the cost of running it lower than its full size suggests.
What does a 1-million-token context window mean?
It means the model can take in a very large amount of text at once — roughly a whole library of documents in a single request — instead of forgetting the start of a long conversation.
Is the one-twentieth cost claim independently checked?
No. It is a vendor-reported figure from Tencent's own internal WorkIndex benchmark, not an independent evaluation, so it should be read as Tencent's own claim.
👁
Source: TechNode