LLMs

Hy4 Flash vs Qwen4-Max: Open Weights or Bigger Model

Two Chinese flagships, five days apart, pointing opposite ways on licensing.

🕐 7 min read 👁 44 views 📅 Nov 21, 2026
The short version
● Same context Both hold one million tokens.
● Opposite licence Hy4 flash ships weights. Qwen4-Max is API only.
● Size Qwen4-Max is 2.4T with 71B active. Hy4 flash is 224B with 12B active.
● The decision Whether you need to run it yourself.

China's two flagship releases this month landed five days apart and point in opposite directions. Alibaba shipped Qwen4-Max on 16 November as a closed, API-only model. Tencent published the Hy4 flash weights on 19 November for anyone to download. Both hold a million tokens of context.

Side by side

Tencent Hy4 flashQwen4-Max
Released19 November 202616 November 2026
WeightsPublishedNot released
Parameters~224B~2.4T backbone
Active per token~12B~71B
Context1M tokens1M tokens
InputText and imageText
LicenceTencent Hy LicenseCommercial API terms
PriceYour own hardware$1.50 in / $4.50 out per 1M
VariantsStandard, FP8, baseqwen4-max

Qwen4-Max is roughly ten times the parameter count and six times the active parameters. On raw capability that gap should show. What it cannot do is run inside your own network.

When Hy4 flash wins

Three cases, and they are all about control rather than capability. Data that cannot leave your infrastructure. Workloads large enough that $1.50 per million input tokens compounds into real money. And anything you intend to fine-tune — Tencent shipped a base model alongside the instruction-tuned one, which is what post-training actually requires.

The 12-billion active figure is what makes any of that realistic. A 224B sparse model with 12B active is servable on hardware a team might own; a dense model of that size would not be.

When Qwen4-Max wins

When you want the stronger model and do not want to operate anything. At $1.50 input and $4.50 output per million tokens it is inexpensive for a frontier-class model, and Alibaba carries the serving cost, the uptime and the upgrades.

It also takes text only, where Hy4 flash accepts images. If your pipeline needs image input, that settles it before price does.

What neither side published

Tencent released no benchmark figures and no hardware requirements with the weights. Alibaba published benchmarks but not the weights. Each withheld exactly what the other provided, so a like-for-like quality comparison is not available from either vendor — you would have to run it yourself.

That is less of an obstacle than it sounds. Download Hy4 flash, run your own workload against both, and you will have a better answer in a day than any benchmark table would give you.

FAQ

Is Hy4 flash open source?
The weights are published on Hugging Face under the Tencent Hy License. Read the licence terms before commercial use — a custom licence is not the same as MIT or Apache.
Is Qwen4-Max available as open weights?
No. Alibaba published benchmark results at launch without releasing weights for Max.
Which has the bigger context window?
Neither. Both hold up to one million tokens.
Which is cheaper?
Qwen4-Max costs $1.50 per million input tokens and $4.50 output. Hy4 flash costs whatever your own hardware costs, which is cheaper at volume and more expensive at low usage.
Which accepts images?
Hy4 flash takes text and image input. Qwen4-Max is text.
⚖ Our Verdict

Need it inside your own network, or intend to fine-tune it? Hy4 flash, which ships a base model too. Want the stronger model with nothing to operate? Qwen4-Max at $1.50 in and $4.50 out.