Share𝕏inWA

Tencent Open-Sources Hy4 Flash: 224B AI Model, 1M Context

TechorCoffee
TechorCoffee
Contributing Writer

Published (Updated )

Tencent Open-Sources Hy4 Flash: 224B AI Model, 1M Context

Quick answer

From hosted API to open weights in two days

Tencent's Hy team, formerly known as Hunyuan, launched Hy4 Flash on November 16, 2026, as a hosted model. On that day, it was available through Tencent Cloud TokenHub and on OpenRouter, but the weights β€” the files that let anyone download and run the model themselves β€” were not public. Tencent said at the time that it planned to open-source the model and iterate on it often.

Two days later, on November 18, the open-source release arrived. The weights are now downloadable, and developers can run the model on their own hardware instead of paying for API calls. That two-step pattern β€” hosted first, open weights second β€” is one Tencent has used before with the Hy family, and it is also the pattern followed by other Chinese AI labs, including the teams behind Qwen and DeepSeek.

The timing matters for a simple reason. A hosted API is useful, but it leaves the model in Tencent's hands: the company sets the price, controls the service and can change the terms. Open weights change the relationship. Once the files are public, developers can run the model locally, fine-tune it for their own tasks, and keep using it even if the hosted service changes.

The specs: big model, small active footprint

Close-up of a modern AI processor chip on a circuit board with glowing blue traces

Hy4 Flash is built as a mixture-of-experts model, which means it has a large total size but only uses a small slice of that size for each token it generates. The headline numbers, as published by Tencent, are:

The 12-billion-active figure is the one that matters most for running costs. A model that only activates 12 billion parameters per token is far cheaper to serve than a dense model of the same total size, which would use all 224 billion parameters every time. That is the whole point of the mixture-of-experts design: keep the knowledge of a big model while paying the running cost of a much smaller one.

The 1-million-token context is the other headline number. A context window that large lets the model read very long documents, whole codebases or extended conversations in a single pass. It puts Hy4 Flash in the same class as the largest context windows available today, including those of the flagship models from Google, OpenAI and Anthropic.

The Engram parameters need a note of caution. Tencent describes 64 billion of the model's parameters as Engram parameters, but the release gives no module or entry count, so the figure is best read as a size rather than a detailed architectural claim. Treat it as Tencent's description of the design, not as an independently confirmed specification.

How cheap is it, really

Hy4 Flash is positioned as the cost tier of the Hy family, and the pricing shows it. The official list price on Tencent Cloud TokenHub is:

That is roughly one twelfth of the input price of the flagship Hy4, which lists at about $0.83 per million input tokens. A launch discount halves the Hy4 Flash rates through November 30, 2026, and the model is free in the WorkBuddy app until the same date.

On OpenRouter, where the model is listed as tencent/hy4-flash, the price is lower still: $0.035 per million input tokens and $0.1405 per million output tokens, with cached input at $0.0015. Third-party providers sometimes price below the official list, so the OpenRouter figure is a marketplace price rather than Tencent's own list price.

To put those numbers in context, the flagship models from the big US labs typically cost several dollars per million input tokens. Hy4 Flash costs a few cents. For developers building agents, coding tools or document processing that burn through millions of tokens a day, that difference is the difference between a product that makes money and one that does not.

The catch, as always with API pricing, is that the cheapest price is not the whole story. Speed, reliability, output quality and the cost of the infrastructure around the model all matter. A model that is cheap per token but needs three attempts to get the right answer is not cheap at all.

Where you can use it today

Software developer at a desk with monitors showing code and an AI chatbot interface

Hy4 Flash is available in more places than most new models manage in their first week. The hosted version runs on:

The open-source release adds the self-hosted route. Developers can download the weights from GitHub, Hugging Face, ModelScope and GitCode, with community support for the GGUF format, which is the format used by popular local-running tools. The three released versions are Hy4 Flash, Hy4 Flash-FP8 and Hy4 Flash Base, giving developers a choice between the standard model, a quantised version that uses less memory, and a base model for fine-tuning.

One detail worth noting for readers in Europe: as of November 17, OpenRouter marked both tencent/hy4 and tencent/hy4-flash as not available through EU in-region routing. That may change, but it is worth checking before building a product that depends on it.

How it performs: what the numbers say, and what they do not

Tencent's own claims for Hy4 Flash are strong: over 150 tokens per second of generation, a KV cache far smaller than its predecessors, and a model distilled from the flagship Hy4 and trained further with reinforcement learning. Distillation means the smaller model was taught by the larger one, which is how it keeps much of the flagship's ability at a fraction of the size.

Independent evidence is thinner, as it usually is in the first days after a release. One early independent test came from Atomic Agent, which ran Hy4 Flash on a coding task through OpenRouter on November 17. The model passed all 59 hidden tests in 4 of 6 runs, at a cost of about $0.02 per task. That is a small, single-task test, not a broad benchmark, but it suggests the model is usable for real coding work at very low cost.

The broader benchmark picture is vendor-reported, which means the numbers come from Tencent itself and have not been independently verified. That is normal for a new model release, and it is also a reason to be careful. Vendor benchmarks are chosen to show the model at its best. The real test will come over the coming weeks, as independent evaluators run the model through standard benchmarks and as developers report how it behaves on their own workloads.

The honest summary is this: Hy4 Flash looks fast and cheap, with a very large context window, and the first independent coding test is encouraging. But the model is days old, the benchmarks are mostly Tencent's own, and nobody outside the company has had time to test it thoroughly yet.

The Hy4 family: where Flash fits

Hy4 Flash is the second model in a family that Tencent has been building through 2026. The flagship, Hy4, launched on November 4, 2026, with 770 billion total parameters and 49 billion active per token, plus image understanding and computer-use features. Tencent says Hy4 ranks second among Chinese models on the Artificial Analysis Intelligence Index and first among open-source models on the Agent Arena leaderboard. Those are vendor-cited rankings, not independent results, but they show where Tencent thinks its flagship stands.

The family has a clear structure:

The idea is the same one Google, OpenAI and Anthropic all use: a big model for the hardest jobs and a smaller, cheaper model for the routine ones. The difference is that Tencent is making both tiers available as open weights, which the US labs generally do not do with their flagships.

The Hy family itself has a longer history. It began as Hunyuan, Tencent's in-house AI effort, and has gone through several generations. Hy3 came before Hy4, and the flash tier's KV cache is 24 times smaller than Hy3's, which shows how much of the engineering work has gone into making the models cheaper to run, not just more capable.

Why this release matters

Open-weight models matter for a different reason than hosted APIs. When the weights are public, the model cannot be taken away, repriced or quietly changed. Developers can run it on their own servers, in their own countries, under their own rules. That is why open releases from Chinese labs have become one of the most watched parts of the AI industry: they give developers everywhere a powerful model that no single company controls.

Hy4 Flash also matters because of where it sits in the market. It is a 224-billion-parameter model with a 1-million-token context that costs a few cents per million tokens. A year ago, that combination did not exist at any price. Today it is downloadable for free. That is a measure of how fast the cost of AI capability is falling, and it is good news for anyone building products on top of these models.

The release also sharpens the competition. Alibaba's Qwen models, DeepSeek's open models and now Tencent's Hy family are all competing to be the default open model for developers. Each release pushes the others to be cheaper, faster or more capable. The winners of that competition are the developers who get better models for less money, and the users who get better products built on them.

There is a geopolitical layer too, which is worth stating plainly. The most capable open models now come largely from Chinese labs, while the most capable closed models come largely from US labs. That split shapes everything from export controls to the question of which models developers in different countries can use. Tencent's decision to open-source Hy4 Flash is part of that pattern, and it is one reason the release is being watched well beyond the developer community.

What we still do not know

A lot about Hy4 Flash is still unanswered, and honest coverage should say so.

The bottom line

Tencent's Hy4 Flash, open-sourced on November 18, 2026, is a 224-billion-parameter AI model with a 1-million-token context that anyone can download and run for free. It is the cheap, fast tier of the Hy4 family, distilled from the 770-billion-parameter flagship, and it costs about $0.07 per million input tokens through the hosted API β€” roughly a twelfth of the flagship's price. The benchmarks so far are mostly Tencent's own, and the model is days old, so independent testing is still to come. But the release is real, the weights are public, and for developers who need a capable model at a very low price, it is one of the most interesting open releases of the year.

FAQs

What is Tencent Hy4 Flash?

Hy4 Flash is the fast, low-cost model in Tencent's Hy4 family. It is a mixture-of-experts model with 224 billion total parameters, 12 billion active per token, and a context window of up to 1 million tokens.

Is Hy4 Flash really open source?

Yes. Tencent open-sourced the model on November 18, 2026, releasing three versions β€” Hy4 Flash, Hy4 Flash-FP8 and Hy4 Flash Base β€” downloadable from GitHub, Hugging Face, ModelScope and GitCode.

How much does Hy4 Flash cost to use?

Through Tencent Cloud TokenHub it lists at about $0.07 per million input tokens and $0.28 per million output tokens, with a 50% discount through November 30, 2026. It is free in the WorkBuddy app until November 30, and OpenRouter lists it at $0.035 per million input tokens.

What is the difference between Hy4 and Hy4 Flash?

Hy4 is the flagship, with 770 billion total parameters and 49 billion active per token. Hy4 Flash is the smaller, cheaper tier, with 224 billion total parameters and 12 billion active per token, designed for everyday work at high speed and low cost.

Where can I try Hy4 Flash?

You can use it through Tencent Cloud TokenHub, OpenRouter as `tencent/hy4-flash`, or in Tencent's WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis apps. Developers can also download the weights and run it themselves.

Can I run Hy4 Flash on my own computer?

You can download the weights, but a 224-billion-parameter model needs powerful hardware to run locally, even in the smaller FP8 version. Most people will find the hosted API or a cloud deployment far more practical.

How good is Hy4 Flash at coding?

Early signs are encouraging: one independent test on November 17 passed all 59 hidden coding tests in 4 of 6 runs at about $0.02 per task. Broader benchmarks are mostly vendor-reported so far, so independent results are still to come.

Was this article helpful?

TechorCoffee

Written By

TechorCoffee

Contributing Writer

TechorCoffee is a technology reporter and software creator writing reviews and tutorials for TechOrCoffee.

Comments (0)

Sort by:

Leave a Comment

Loading comments...

Read Next