Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Tencent released downloadable weights for its flagship Hy4 model on November 20, with a newly linked System Card. The weights are governed by Tencent Hy License 1.0, not Apache 2.0, and the license adds commercial conditions for some operators, according to a report from Pandaily.
The release follows Hy4's November 4 service launch. It is separate from the smaller Hy4 flash checkpoints made available earlier this week.
Model specifications
The official model card describes a multimodal mixture-of-experts system:
- Language backbone: 770 billion parameters
- Active parameters: 49 billion per text token
- Context window: 1,048,576 tokens
- Not counted in the backbone figure: the vision encoder, the adapter and the DFly2 drafting component
Tencent lists three checkpoints: instruction-tuned Hy4, Hy4-FP8 and pretrained Hy4 base. These give developers different starting points for deployment or further training.
What was verified
Pandaily confirmed anonymous retrieval of a small byte range from a weight shard in each of the three repositories. It did not download complete checkpoints or run inference, so performance and behavior of the released weights remain untested by the outlet.
The deployment documentation includes vLLM and SGLang recipes that use eight-way tensor parallelism. Pandaily describes these as substantial infrastructure deployments, not evidence that the flagship runs on an ordinary laptop.
License terms
The additional commercial conditions in Tencent Hy License 1.0 apply to operators that provide third parties with general-purpose remote model inference or fine-tuning (model-as-a-service). Those operators need separate written permission from Tencent in either of two cases:
- Aggregate worldwide gross revenue exceeds $50 million in any consecutive 12-month period, regardless of whether Hy4 generated that revenue.
- The model is used in a product or service exceeding 100 million monthly active users in any calendar month.
The license excludes feature-limited end-user workflows and simple relays to independently operated model services from its definition of model-as-a-service. A narrow internal-use exception applies if neither the model, its derivatives nor their outputs are made available to third parties. Pandaily notes the thresholds should not be presented as a blanket restriction on every commercial user.
Security claims
Tencent says it filtered offensive cybersecurity training material and ran internal specialist security audits. These are the company's own assessments, not independent assurances that the model cannot be misused. Developers still need their own evaluation of task performance, resource requirements and security controls.
Benchmark scores and per-token pricing were not reported in the article.
What this means
The change is access to flagship model artifacts and the terms that govern them. Hy4's previously announced vision and computer-use capabilities are not a new debut here.
For teams that fine-tune or serve the model for their own products, the main questions are infrastructure and license scope. Eight-way tensor parallelism in the vLLM and SGLang recipes puts self-hosting in the multi-accelerator range. The 770B backbone figure excludes the vision encoder, adapter and DFly2 component, so total deployment footprint should be checked against the model card.
For companies that resell inference or fine-tuning, the license is the deciding factor. The $50 million revenue threshold counts all worldwide gross revenue, not only revenue tied to Hy4. Operators near that line, or building on a product with a large user base, should read the model-as-a-service definition and its exclusions before committing.
The availability of three checkpoints, including the pretrained base, means developers can choose between using the instruction-tuned model directly, running the FP8 variant, or training from base. Tencent's security statements are unverified by third parties, so independent evaluation remains the developer's job.
Related Articles
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Tencent releases Hy4 flash weights in three variants under a custom license with revenue and user caps
Tencent has released downloadable weights for Hy4 flash in three variants: instruction-tuned, FP8 and base. The checkpoints ship under Tencent Hy License 1.0, which requires a separate license above US$50 million in revenue or 100 million monthly active users. Tencent says the larger Hy4 model's weights are still being prepared.
Tencent releases Hy4 Flash Base: 220B-parameter MoE, 12B active, 1M-token context
Tencent has published Hy4 Flash Base on Hugging Face: a 220B-parameter mixture-of-experts model with 12B active parameters per token, text and image input, and a 1M-token context. Tencent calls it its most cost-effective Hy model to date.
Tencent Hunyuan says Hy4 flash processed over 50 trillion tokens on its first day
Tencent's Hunyuan says Hy4 flash processed more than 50 trillion tokens on its first day after launching on November 16. The company says it is the highest first-day usage among models in the Hunyuan series.
Comments
Loading...