The Pulse
Tencent releases Hy4 weights with limits for large service providers
Tencent has made its 770-billion-parameter Hy4 model available as open weights, alongside FP8 and base versions. Large model-as-a-service businesses face separate licensing requirements under Tencent's terms.

Tencent has released downloadable weights for Hy4, a multimodal AI model whose language backbone has 770 billion parameters. The release became public on November 20, Sulat AI & Tech News reported, with the main model, an FP8 version and a pretrained base model available on Hugging Face, ModelScope and GitCode.
Hy4 uses a mixture-of-experts architecture, activating 49 billion language parameters per token. It accepts text, images and video, and its context window holds 1,048,576 tokens, according to Tencent's model card.
Three sets of weights
The release includes the instruction-tuned Hy4, Hy4-FP8 and Hy4 base. Sulat reports that the FP8 version stores the model in 8-bit numbers, while the base version is the pretrained model before chat tuning.
The files total about 1.55 TB for Hy4, 0.79 TB for Hy4-FP8 and 1.54 TB for the base model on Hugging Face, according to the outlet. A system card covering the model's safety evaluation accompanied the weights.
The language backbone has 78 layers: one dense layer and 77 mixture-of-experts layers. Each mixture-of-experts layer contains 256 routed experts and one shared expert; a token activates eight routed experts as well as the shared one.
A separate vision encoder has 1.23 billion parameters. The model also includes a 3.3-billion-parameter DFly2 drafter for speculative decoding, a technique in which a smaller model proposes tokens for the larger model to check.
License conditions
Hy4 uses the Tencent Hy License, rather than the Apache 2.0 license used for Hy4 preview. Sulat reports that the new license is derived from Apache 2.0 and adds commercial terms for large model-as-a-service businesses.
A business providing third parties remote access to inference or fine-tuning needs a separate written license if its worldwide gross revenue exceeds US$50 million in any 12-month period, according to Sulat. The outlet reports that a separate license is also needed before such a business uses the model in a product with more than 100 million monthly users.
Sulat reports that internal use is exempt provided neither the model, a derivative nor their outputs reach third parties. Fine-tuned, quantized, distilled, merged and pruned versions count as derivative works under the terms, the outlet says.
Evaluation and limitations
Tencent's model card describes uses including computer workflows, game development, financial work and research. In its evaluation involving 213 finance professionals and 579 financial tasks, Hy4 received an average score of 3.04 out of 4, compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3.
The company says Hy4 still trails top proprietary models on some difficult internal tests, including research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. Its base-model benchmark table also carries a warning that exposure to synthetic instruction data during mid-training can sometimes cause the model to miss the expected answer format in few-shot evaluations, potentially distorting scores.
The release supports deployment with the vLLM and SGLang inference frameworks, Sulat reports. The outlet says the repository also includes fine-tuning scripts for DeepSpeed, LLaMA-Factory and ms-swift.
Sources
- Original source Tencent · page 1 huggingface.co
Other pages read
- Tencent · page 2 huggingface.co
- Sulat AI & Tech News news.sulat.com
In the tracker
Related stories
- Endless Frontier releases weights for 180B bioinformatics model
- Tencent releases Hy4 flash with lower API prices than Hy4
- Tencent Brings Hunyuan Hy4 to Cloud ADP With Visual Understanding and Computer Operation
- Kyutai releases AURA-5B for non-commercial talking-head video
- Tencent Says Agents That Probed Amap Came From Third-Party Model Evaluations