Short answer: Hy4 Flash is Tencent's cheaper open model in the Hy family, released on November 16, 2026. It's a mixture of experts model that accepts text and images, writes text, and supports a context window of up to 1 million tokens. Tencent's announcement lists 224 billion total parameters, including 64 billion Engram parameters, with only 12 billion active. The Hugging Face model card lists 220 billion total and the same 12 billion active figure (11.3 billion for text plus 0.43 billion for vision). On Tencent Cloud TokenHub, the list price is $0.07 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens, checked on November 20, 2026.
If you saw "Hy4 Flash" or hy4-flash in a model picker and wanted the cheap option, this guide covers what it is, what it costs, how to call it, what the license allows, and when a bigger Hy model or another open model is the better pick.
What is Hy4 Flash?
Hy4 Flash is the low cost member of Tencent's Hy model family. Tencent says it is the most cost effective Hy model to date, and it recommends Flash for everyday work: research, documents, slides, simple front end displays and computer use. For long, hard software engineering, it points you to the larger Hy4 model instead (Tencent).
A mixture of experts model keeps a large set of specialist networks, but only turns on a few of them for each token. That's why the active count (about 12 billion) is much smaller than the total. The card also describes a built in drafter, DFly2, used for speculative decoding, which is a way to speed up generation.
Tencent says decoding can exceed 150 tokens per second when Flash is served with its own optimizations (megakernel, 4-bit quantization and DFly). That's a vendor figure, not an independent lab test, so treat it as a claim to check on your hardware.
How much does it cost?
Tencent's announcement prices Hy4 Flash at about $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens. Tencent Cloud's TokenHub price list, checked on November 20, 2026, matches that list price under the model id hy4-flash (Tencent Cloud TokenHub).
Two limited offers run until November 30, 2026:
- WorkBuddy: Tencent says you can use Hy4 Flash for free until that date.
- TokenHub: calls are billed at 50% of the list price until 12:00 Beijing time on November 30, 2026. After that, the list price applies.
Tencent also says Flash is available in Yuanbao, WeCom, QQ, ima and Marvis, and that OpenRouter integration is coming soon. There's no date for OpenRouter yet.
Those are Tencent's prices, dated November 20, 2026. If you self host the weights, you pay for your own compute instead.
How do you run it?
The weights live on Hugging Face as tencent/Hy4-flash, with an FP8 variant at tencent/Hy4-flash-FP8. Tencent also published recipes for an OpenAI compatible API and for production serving with vLLM or SGLang.
If your stack already talks to an OpenAI style endpoint, you can point it at a local or TokenHub server and use the model name hy4-flash. For serving, Tencent's card shows official images, vllm/vllm-openai:hy4-flash and lmsysorg/sglang:hy4-flash, typically with the FP8 checkpoint and tensor parallelism.
You don't have to self host. TokenHub is the hosted path if you just want an API.
What does the license allow?
Hy4 Flash is released under the Tencent Hy License, Version 1.0, not Apache 2.0 or MIT. The license text is derived from Apache 2.0, but Tencent added commercial terms in Section 10, and it says the license is not affiliated with the Apache Software Foundation.
You can use, modify and redistribute the weights, including for most products, as long as you keep the license notice. The extra rule is about Model as a Service, meaning you give third parties remote access to inference or fine tuning, for example through an API. In that case you need a separate written license from Tencent if either of these is true:
- your worldwide gross revenue exceeds $50 million in any consecutive 12 month period, or
- the product or service has more than 100 million monthly active users in any calendar month.
That extra rule does not apply to internal use, as long as the model, derivatives and their outputs aren't made available to third parties. It also doesn't apply to ordinary end user products whose model features stay inside a specific workflow, or to apps that only relay requests to someone else's API. For licensing questions, Tencent lists hy_partners@tencent.com.
How does it compare with Qwen4-Max and Edge0-145B?
All three are options when you want an open or open weight model instead of a closed API, but they solve different problems.
- Qwen4-Max is Alibaba's flagship. Reach for it when you want the strongest model in that family and you're fine paying flagship prices.
- Edge0-145B is built to run on your own computer. Reach for it when you want the weights on hardware you control, and you're ready to deal with setup and VRAM.
- Hy4 Flash is the cheap hosted option with open weights, a 1 million token window, and image input. Reach for it when you want a low token bill or a model you can also serve yourself, and the task is everyday work rather than a long software engineering run.
Tencent's own guidance is the same split: Flash for daily tasks, Hy4 when the job is a long coding project.
Go deeper
- What is Qwen4-Max: Alibaba's flagship model, how to call it, and what it costs.
- What is Edge0-145B: the open 145B model built to run on your own machine.
- AI tools for developers: the 2026 tools, sorted by what they do.
Picking a cheap open model is a skill
New open models now show up every few weeks, each with its own license and price list. Knowing how to read a model card, check a license and test the model on your own tasks will outlast any single release. If you want to get better at building with models and agents, at your own pace, see how the AI Engineering program works.
Sources, checked November 20, 2026: Tencent's Hy4 Flash announcement (November 16, 2026), the tencent/Hy4-flash model card and Tencent Hy License, and the Tencent Cloud TokenHub price list.
