wasl.tech

Tencent Hy4 Flash: 224B-Parameter AI Matches Claude Haiku 5.5 at a Fraction of the Cost

Picture a small business owner at nine in the evening, staring at a pile of supplier quotes, a half-finished slide deck and a spreadsheet that refuses to balance. Hiring help costs money. The big AI models cost money too — or so she thought.

On November 16, Tencent quietly changed that maths. Its new Hy4 Flash model handles the everyday office work — the research, the slides, the spreadsheets — for roughly the price of a coffee per million words. And according to Tencent, users like it almost as much as Anthropic’s Claude Haiku 5.5.

The launch barely made a ripple outside China. That is a mistake. Hy4 Flash is the third release in Tencent’s Hy4 family in less than three months, and it shows just how fast the price of capable AI is falling.

What Tencent Hy4 Flash Actually Is

Hy4 Flash is the budget member of Tencent’s Hunyuan model family. It is not a cut-down demo or a toy model. It is a full mixture-of-experts system with 224 billion parameters in total, of which only 12 billion switch on for any single piece of work.

That design is the whole point. A model that wakes up a small slice of itself for each task can answer faster and cheaper than one that fires everything at once. Tencent built Hy4 Flash for the unglamorous jobs — research, slide-building, spreadsheet edits — rather than the hardest reasoning problems.

The model accepts text and images, understands documents and charts, and can call tools. It cannot generate images or video. Its output is text, delivered through an API or inside Tencent’s own apps.

Tencent positions it below the flagship Hy4, which arrived on November 4 with 770 billion parameters. The preview of that flagship shipped back in August. Flash arrived just twelve days after the full model — a release pace that says a lot about the competition inside China’s AI labs.

The Numbers Behind Hy4 Flash

The headline figures are striking. Hy4 Flash carries a context window of one million tokens, so it can hold an entire company’s worth of documents in a single conversation. Its maximum input is 960,000 tokens, and it can produce up to 64,000 tokens in one reply.

Under the hood, Tencent lists three ingredients: hybrid attention, a technique called LatentMoE, and a memory system named Engram. Of the 224 billion parameters, 64 billion belong to that Engram memory. The company has published few details about how either works.

Then comes the price. Through Tencent Cloud’s TokenHub, Hy4 Flash costs $0.07 per million input tokens and $0.281 per million output tokens, with cached input at $0.003 per million. A 50 percent launch discount runs through November 30, cutting those rates in half.

Put another way, the flagship Hy4 costs roughly twelve times as much per input token. For a business running thousands of routine AI jobs a day, that gap is the difference between an experiment and a budget line.

Matching Claude Haiku 5.5 — At a Fraction of the Price

Tencent’s boldest claim is about quality, not cost. In a blind test, 85 users of its WorkBuddy agent rated outputs across 259 everyday tasks. Hy4 Flash scored 2.79 out of 4 on satisfaction — close to Claude Haiku 5.5, according to Crypto Briefing.

The median task took 6.7 minutes and cost about 0.26 yuan, roughly four US cents. On Tencent’s internal WorkIndex benchmark, the model completed tasks at one-twentieth the cost of the full Hy4 model.

These are Tencent’s own measurements, and they should be read that way. No independent lab has yet verified the satisfaction scores. But the direction of travel is hard to miss: a model aimed squarely at Anthropic’s cheapest tier, at a price its rivals will struggle to match.

The comparison matters because Claude Haiku 5.5 has become the default choice for high-volume, low-cost work. If a Chinese model can match it on the tasks people actually run all day, the market for budget AI gets a serious new competitor.

Where You Can Use Hy4 Flash Today

Hy4 Flash is already live in several places. It runs inside WorkBuddy, Tencent’s AI agent for office work, where access is free until November 30. It is also available in Yuanbao, Tencent’s consumer chatbot, plus WeCom, QQ, the ima app and Marvis.

Developers can reach it through Tencent Cloud’s TokenHub under the API identifier hy4-flash. The API guide supports the OpenAI Chat Completions and Responses protocols, as well as Anthropic’s Messages format, so existing code should need little change.

An OpenRouter listing has also appeared, giving the model a route to developers outside Tencent’s own cloud. Tencent says it will release the model weights, letting developers download and run Hy4 Flash themselves instead of paying per call.

The Open-Weights Promise — With a Catch

That last promise deserves a closer look. Tencent says it will publish the weights, but it has not named a date or a licence. The Hy4 preview shipped under the plain Apache 2.0 licence, which allows commercial use with few strings attached.

Whether those terms carry over to Flash is unknown. Until the licence file is published, developers should not assume they can use the model freely. Other Chinese releases this year have come with territorial restrictions and revenue limits — MiniMax’s H3-Image, for example, bars use in the United States and Europe.

The weights are also not part of today’s launch. For now, Hy4 Flash is a hosted service only. Anyone planning to self-host will have to wait, and the wait may be long.

Why This Launch Matters

Hy4 Flash lands in the middle of a fierce price war among Chinese AI labs. Alibaba moved its flagship Qwen4-Max into general availability on the very same day, at $1.50 per million input tokens. Meituan’s LongCat-2.5 went open-weight under the MIT licence earlier in November.

Each lab is carving out a different corner. Alibaba is chasing the top of the benchmark tables. Meituan is betting on openness. Tencent is betting that the biggest market is not the hardest tasks, but the most common ones — the daily grind of documents, decks and data.

That bet makes sense. Most companies do not need a model that can prove a theorem. They need one that can summarise a contract, draft a quote and fix a spreadsheet, thousands of times a day, without a scary bill at the end of the month.

The launch also shows how quickly Tencent is moving. Three Hy4 releases in under three months, plus the Hy Realtime voice model that sings and searches the web, suggest a lab shipping at full speed rather than polishing a single flagship.

What Happens Next

The first test will be whether developers outside China pick it up. The OpenRouter listing helps, and the price is hard to ignore. But trust is earned slowly, and Tencent’s benchmark claims will need independent checks before large companies build on them.

The second test is the weights. If Tencent publishes them under a permissive licence, Hy4 Flash could become a staple of self-hosted AI, the way open models from other Chinese labs have. If the licence comes with restrictions, the model will stay a cloud service with a ceiling on its reach.

Either way, the message of November 16 is clear. Capable AI is getting cheaper, faster than most people expected.

The small business owner staring at her spreadsheet at nine in the evening may soon find that the help she needs costs less than the coffee she is drinking. That is the real story of Hy4 Flash — not a new benchmark record, but a new price floor.

Source: Crypto Briefing

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top