Skip to content

Tencent Hy4 open weights: specs, license and how to run.

Tencent released open weights for Hy4, its 770B-parameter multimodal MoE flagship with a 1M context. Specs, benchmarks, license terms and how to run it.

By The Model Atlas · 7 sources

In this post9 sections
  1. 01What did Tencent release?
  2. 02Hy4 specs at a glance
  3. 03What’s new compared with Hy4 preview
  4. 04Hy4 benchmarks
  5. 05Hy4 license: open, with a big-company clause
  6. 06Safety: the Hy4 System Card
  7. 07How to run Hy4
  8. 08What people are building with Hy4
  9. 09Who is Hy4 for?

Tencent has released the open weights for Hy4, its flagship large language model. The Tencent Hy team announced the Hy4 open weights on 20 November 2026, sixteen days after Hunyuan Hy4 first launched through Tencent’s apps and API. Anyone can now download a 770-billion-parameter multimodal mixture-of-experts model with a 1-million-token context. Here is what was released, what Tencent says the model can do, what the license allows and how to run it.

What did Tencent release?

According to the Hy4 model card (opens in a new tab), Tencent has open-sourced three checkpoints on Hugging Face, ModelScope and GitCode:

  • Hy4: the instruction-tuned model.
  • Hy4-FP8: an FP8-quantized version of the instruct model.
  • Hy4 base: the pre-trained base model. Tencent notes that it is not an instruction-tuned chat model.

Tencent also links a GGUF build for testing, a GitHub repo, a finetuning guide and the Hy4 System Card. When Tencent launched Hy4 on 4 November, its announcement said model weights and system cards would follow “in the coming days”. This release delivers on that. If you want the cheaper sibling, see our Hy4 flash explainer.

Hy4 specs at a glance

The model card lists these specifications:

  • Mixture-of-Experts: 770B total parameters in the language backbone, with 49B active per token (plus 1.23B for vision).
  • 78 layers: one dense layer and 77 MoE layers. Each MoE layer has 256 routed experts and 1 shared expert, and every token uses the top 8 routed experts plus the shared one.
  • Context window: 1M tokens (1,048,576).
  • Multimodal input: text, image and video understanding, through a 1.23B-parameter vision encoder that supports dynamic image resolutions.
  • Attention: Gated DeepSeek Sparse Attention with IndexCache, which Tencent says was inspired by DeepSeek and GLM. The indexer picks the top 2,048 tokens of context for each step.
  • Built-in drafter: a 5-layer, 3.3B-parameter model called DFly2 for speculative decoding. Tencent says it reuses Hy4’s KV cache and sparse indices, so each draft token attends to at most about 2K keys whatever the context length, with no extra KV memory.

What’s new compared with Hy4 preview

Tencent says Hy4 substantially scales up both pre-training and post-training compared with Hy4 preview. It adds vision for the first time and brings major gains in instruction following, long-context understanding, long-horizon tasks and deep research. Tencent’s 4 November announcement also said Hy4 introduced computer use for the first time, working across command-line and graphical interfaces, with specialized training on software such as Blender and CAD.

The model card highlights a few use cases with Tencent’s own numbers:

  • Game development: Tencent says it worked with developers at Tencent Interactive Entertainment Group and adapted Hy4 for Unreal Engine, Unity and Godot. It reports gains of 19.1% on GameDev and 9.8% on GameLogic over Hy4 preview.
  • Finance: across 579 real-world financial tasks rated by finance professionals, Tencent reports an average score of 3.04 out of 4 for Hy4, versus 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3, ranking it first among the open-weight models compared.
  • Fewer hallucinations: in internal evaluations, Tencent says Hy4 cut the hallucination rate from 15.76% on Hy4 preview to 5.45%, and speculative errors from 23.1% to 5.6%.

Hy4 benchmarks

Most of Tencent’s published comparisons are charts, so we only quote the numbers it lists as text. In few-shot tests of the pre-trained models, Tencent reports that Hy4 base scores 91.6 on MMLU, 73.9 on MMLU-Pro, 62.7 on SuperGPQA and 57.2 on SimpleQA. In the same table, DeepSeek-V4-Pro-Base scores 91.1, 73.4, 54.8 and 52.1. Hy4 base does not lead every row: DeepSeek-V4-Pro-Base is ahead on MBPP-plus code (79.1 vs 74.0) and OCW Courses math (68.2 vs 66.7).

Tencent adds a caveat: its base model saw a lot of synthetic instruction data in mid-training, so it sometimes ignores the expected answer format in few-shot tests, which may distort scores. These are Tencent’s own results, not independent evaluations.

Known limitations

Tencent is upfront about the gap with closed models. It says that although Hy4 is one of the strongest open-weight models available, a notable gap remains between it and top proprietary models, particularly on hard internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.

Hy4 license: open, with a big-company clause

Hy4 is released under the Tencent Hy License, not Apache 2.0. The license text says it is derived from Apache License 2.0 with modifications, mainly additional commercial terms for large Model-as-a-Service (MaaS) businesses in Section 10, and that it is not the Apache License.

Under Section 10, a company offering third parties general-purpose remote access to the model, for example through an API, must get a separate written license from Tencent if its worldwide gross revenue tops US$50 million in any 12-month period, or before using Hy4 in a product or service with more than 100 million monthly active users. Purely internal use, where neither the model nor its outputs reach third parties, is exempt.

Safety: the Hy4 System Card

Tencent published the Hy4 System Card with the weights. Because safeguards on its hosted API may not carry over once weights are public, Tencent says it built safety into training itself. The report covers frontier risks (with a focus on cybersecurity and biology), content safety, agentic security and misalignment.

How to run Hy4

There are two routes: self-host the open weights, or call a hosted API.

  1. Download the weights from Hugging Face (opens in a new tab) (or ModelScope or GitCode). Pick Hy4, Hy4-FP8 or Hy4 base.
  2. Serve it with vLLM or SGLang. Tencent recommends both for production and provides prebuilt images (`vllm/vllm-openai:hy4` and `lmsysorg/sglang:hy4`). Its example commands serve Hy4-FP8 with tensor parallelism across 8 GPUs and turn on DFly2 speculative decoding.
  3. Call the OpenAI-compatible endpoint your server exposes. Tencent recommends temperature 0.9 and top_p 1.0. Reasoning defaults to `high`; pass `no_think` for direct answers.

Most people will start with a hosted API. Tencent’s 4 November announcement says Hy4 is available in Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima and more, and through third-party platforms including OpenRouter. At the time of writing, OpenRouter’s Hy4 listing shows a 1,048,576-token context and about $0.42 per million input tokens and $1.25 per million output tokens.

What people are building with Hy4

Builders have been using Hy4 since its 4 November launch, mostly for games and interactive web pages, which lines up with Tencent’s game-development focus. A few highlights from The Model Atlas:

See every project on the Hunyuan Hy4 model page, and compare with Hy4 flash builds.

Who is Hy4 for?

Going by Tencent’s description, the open weights suit teams that want a strong multimodal model they can host and fine-tune themselves, especially for game development, frontend, finance and office work, research and computer-use agents. Large MaaS providers above the license thresholds need a separate deal with Tencent.

When were the Hy4 open weights released?

Tencent Hy announced the open weights on 20 November 2026. Hy4 itself launched on 4 November 2026 through Tencent’s products and API.

How big is Hy4?

Tencent’s model card lists 770B total parameters with 49B active per token, a 1.23B vision encoder, a 3.3B DFly2 drafter and a 1M-token context window.

Is Hy4 Apache 2.0?

No. It uses the Tencent Hy License, which is derived from Apache 2.0 but adds commercial terms: MaaS businesses with over US$50 million in annual revenue, or products with over 100 million monthly active users, need a separate license from Tencent.

Can I run Hy4 on a laptop?

Hy4 is built for server hardware: Tencent’s own serving examples run Hy4-FP8 across 8 GPUs with vLLM or SGLang. Tencent also links a GGUF build for testing, but a hosted API such as OpenRouter or Tencent Cloud TokenHub is the easiest way to try it.

How is Hy4 different from Hy4 flash?

Hy4 is the flagship. Hy4 flash, launched on 16 November, is a smaller and cheaper sibling that Tencent recommends for most everyday tasks.

Sources · 7

  1. tencent/Hy4 model card (Hugging Face) (opens in a new tab)
  2. Tencent Hy License (Hy4 LICENSE file, Hugging Face) (opens in a new tab)
  3. Hy4 System Card (Tencent Hy, November 2026) (opens in a new tab)
  4. Tencent Launches Hy4, Joining the Ranks of the World’s Leading Open-Source AI Models (Tencent, 4 November 2026) (opens in a new tab)
  5. Tencent Launches Hy4 flash to Power High-Speed, Cost-Efficient AI Deployment for Everyday Workflows (Tencent, 16 November 2026) (opens in a new tab)
  6. Tencent-Hunyuan/Hy4 (GitHub) (opens in a new tab)
  7. Tencent: Hy4 (OpenRouter) (opens in a new tab)
All 5 builds →