NEWS

Tencent releases Hy4 Flash weights for local deployment

Tencent releases Hy4 Flash weights for local deployment
Image: Redação iMasters
News
Reading
3 min
Tables
1
In short

Two days after launching Hy4 Flash via app and cloud API, Tencent published the model's weights for download, including an FP8 version and a base model for further training.

Tencent published the weights of its Hy4 Flash model on Wednesday (November 18, 2026), two days after launching the version via its own apps and cloud API. The release fulfills what the company had already announced: that it would make the model's weights available for anyone who wanted to run it outside Tencent's infrastructure. The information comes from TechNode.

In addition to the standard model, Tencent published an FP8 version (an 8-bit floating-point format that reduces memory usage and speeds up inference) and a base model, intended for those who want to fine-tune or further train it. All three packages came with deployment material and are released under the Tencent Hy License, the company's own license for its Hy model family.

What Hy4 Flash delivers

The model accepts text and image input and supports a context window of up to one million tokens, enough to process long documents, entire spreadsheets, or large codebases in a single call. Tencent positions Hy4 Flash for everyday productivity tasks: research, document preparation, and spreadsheet work.

SpecificationValue
Total parameters~224 billion
Parameters activated per token~12 billion
Context windowup to 1 million tokens
Supported inputstext and image
Published versionsstandard, FP8, base model

The large gap between total parameters (224 billion) and what gets activated per token (12 billion) is the signature of a mixture-of-experts (MoE) architecture, a pattern that has become common in recent releases of models at this scale precisely because it allows for a large model without paying the computational cost of activating everything on every inference. Tencent doesn't detail the architecture in the published material, but the ratio between the two numbers points in that direction.

Why this matters for those running models locally

For developers in Brazil, the main news isn't the launch itself (which had already happened two days earlier, via app and API), but the fact that the weights are no longer under Tencent's exclusive cloud control. A model with 12 billion parameters activated per token is much lighter to serve than dense models with hundreds of billions, which lowers the hardware bar needed to put something competitive into production without depending on a pay-per-token API.

In practice, this opens up three paths an engineering team can evaluate:

  • Running it on in-house infrastructure, avoiding rate limits, per-call costs, and dependence on a third-party API's availability.
  • Using the FP8 version on hardware with less available VRAM, trading off some numerical precision for memory savings.
  • Starting from the base model for domain-specific fine-tuning, something closed models simply don't allow.

The one-million-token context is also relevant for those building RAG (retrieval-augmented generation) pipelines: with such a large window, it's possible to drop entire documentation sets, code repositories, or knowledge bases directly into the prompt, reducing the chunking complexity that would normally be needed with smaller windows.

What's still unclear

The material released by Tencent includes the weights and deployment documentation, but TechNode's report doesn't detail the specific terms of the Tencent Hy License or provide benchmarks comparing Hy4 Flash with other open models of similar weight, such as those in the Llama, Qwen, or DeepSeek families. Before putting the model into production, it's worth reading the license carefully: proprietary licenses from large Chinese companies often include commercial-use clauses and active-user limits that differ from the Apache or MIT licenses part of the community is used to.

As of the source's publication, there are also no public benchmark performance numbers that allow comparing Hy4 Flash with direct competitors on coding, reasoning, or tool-use tasks. For those evaluating the model, the practical path is to download the weights, run your own load and quality tests on the specific use case, and only then decide between keeping the dependency on Tencent's cloud API or migrating to local deployment.

Translated from the Brazilian Portuguese original · Read the original