Tencent Open-Sources Hy4 flash: 12B of 224B Parameters Active, 1M-Token Context
Tencent has open-sourced the weights of Hy4 flash, and the vLLM engine supported it on day one. The model has 224B parameters in total, but only 12B are active for each token, so responses should be faster and cheaper than from a dense model of the same size. The 1M-token context equals a stack of thick books or a batch of contracts in a single request. A ready-made Docker image is available, so you can test the model on your own documents in an evening, but you will need memory for all 224B parameters.
AI-processed from X: @vllm_project; edited by Hamidun News
Open models increasingly run into a practical question rather than a quality one: what to run them on, and how much each response will cost. Tencent has released Hy4 flash with open weights, and vLLM, a popular engine for serving models, added support on release day. Here is how to read the numbers in the name.
224B is the total parameter count, the model's "volume of knowledge." 12B is how many of those parameters participate in processing each token, so this figure determines the speed and price of a response. 1M tokens is the context size: roughly several thick books or a large batch of contracts that you can load into a single request.
If you host models yourself, you won't have to wait for the community to polish the setup.
What exactly was released
Hy4 flash was built by the Tencent Hunyuan team. According to the description, it is a natively multimodal model: it works not only with text but also with images, documents, and slides. The stated use cases are research tasks, work with large documents and presentations, and computer use, where the model looks at the screen and acts like a user.
Three variants are available on Hugging Face: the main Hy4 flash, the quantized Hy4 flash-FP8, and the base Hy4 flash-Base. FP8 stores weights in 8 bits and takes half the memory of the 16-bit version. The Base version is not tuned for dialogue and is meant for those who want to do their own fine-tuning.
How to understand "224B, but 12B"
This is the mixture-of-experts principle. Picture a large law firm: it employs hundreds of people, but only three work on any given case. The same applies here: the model has many "specialists," and only a small portion is engaged for each token. Quality depends on the total volume of knowledge in the weights, while the speed and price of a response depend on the active part. That is why Hy4 flash should run faster than a dense 224B-parameter model.
Among its architectural choices, Tencent names hybrid attention, Latent MoE, and Engram. Of the 224B parameters, 64B belong to Engram alone. The announcement does not explain how this block works; for details, check the model card. Tencent sums up the overall idea this way: efficiency is placed at the center of the architecture. The model is positioned as fast and built for everyday productivity.
What this means in practice
The main news for engineers is day-zero support in vLLM. A ready-made Docker image, vllm/vllm-openai:hy4-flash, and a deployment recipe at recipes.vllm.ai are available. That means you can spin up an OpenAI-compatible server with the model using the same tools as for other open LLMs, without building anything yourself or hunting for compatible versions.
What to keep in mind:
- Memory. A rough estimate: 224B parameters in FP8 take about 224 GB for the weights alone, and twice that in 16-bit format. Only 12B parameters are active, but all of them must be held in memory. So you will need a server with several GPUs, not a single card.
- Long context is expensive. A million tokens in a request loads the memory used for the attention cache, and you will have to pick the practical length to fit your hardware.
- The announcement gives no quality figures. The original post contains no benchmarks, no comparisons with other open models, and no license terms. Check these before building a product on the model.
Tencent also says Hy4 will be released soon. The flash version looks like the first model in the family available for self-hosting.
Who should take a look
If you process long documents and slides, or are experimenting with agents that look at the screen, Hy4 flash is worth running on your own data. Start with the FP8 version and the vLLM image, measure speed on real requests, and compare quality with the model you use now. Until independent tests appear, it is too early to judge by phrases like "fast and efficient." But the combination of open weights, multimodality, a large context, and a ready-made launch recipe makes the release a notable candidate for experiments.
- Nov 19, 2026Amazon and Google will hold 28.4% of Anthropic after its IPO, but without voting rights
- Nov 19, 2026OpenAI and AMD Find Up to 14% PPA Gain in Flagship GPU Block in Five Months
- Nov 19, 2026Anthropic Heads for the Stock Market: $17 Billion in Quarterly Revenue and $895.4 Billion in Server Contracts
- Nov 19, 2026AI token prices fell 21.2% in six weeks, while corporate API spending dropped 6.3%
Key Questions Answered
What exactly was released?
Hy4 flash was built by the Tencent Hunyuan team. According to the description, it is a natively multimodal model: it works not only with text but also with images, documents, and slides. The stated use cases are research tasks, work with large documents and presentations, and computer use, where the model looks at the screen and acts like a user. Three variants are available on Hugging Face: the main Hy4 flash, the quantized Hy4 flash-FP8, and the base Hy4 flash-Base.
How to understand "224B, but 12B"?
This is the mixture-of-experts principle. Picture a large law firm: it employs hundreds of people, but only three work on any given case. The same applies here: the model has many "specialists," and only a small portion is engaged for each token. Quality depends on the total volume of knowledge in the weights, while the speed and price of a response depend on the active part. That is why Hy4 flash should run faster than a dense 224B-parameter model.
What this means in practice?
The main news for engineers is day-zero support in vLLM. A ready-made Docker image, vllm/vllm-openai:hy4-flash, and a deployment recipe at recipes.vllm.ai are available. That means you can spin up an OpenAI-compatible server with the model using the same tools as for other open LLMs, without building anything yourself or hunting for compatible versions. What to keep in mind: Memory. A rough estimate: 224B parameters in FP8 take about 224 GB for the weights alone, and twice that in 16-bit format.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.