Tencent Releases 770-Billion-Parameter Hy4 Model Weights

In this article (8)
Tencent released downloadable weights for its flagship Hy4 artificial intelligence model on November 20, publishing base, instruction-tuned and 8-bit floating point checkpoints under conditional commercial licensing.
The system runs on a multimodal mixture-of-experts language backbone totaling 770 billion parameters, activating 49 billion parameters per text token across a context window of 1,048,576 tokens.
The parameter total excludes the model’s vision encoder, adapter and DFly2 drafting component. Tencent published deployment documentation specifying vLLM and SGLang recipes that require eight-way tensor parallelism, establishing heavy compute demands for enterprise teams attempting to run the flagship checkpoints on private infrastructure.
Commercial restrictions target high-volume service providers
Tencent released the files under its proprietary Tencent Hy License 1.0 rather than open-source terms such as Apache 2.0. The license imposes commercial restrictions on platforms that offer remote general-purpose inference or fine-tuning services to third parties.
Such model-as-a-service providers must obtain written permission from Tencent if their annual worldwide gross revenue exceeds $50 million across any consecutive 12-month period. That revenue cap applies whether or not the Hy4 architecture directly generated the turnover.
Operators must also seek Tencent approval before serving any product or consumer digital service that exceeds 100 million monthly active users in a single calendar month.
The agreement carves out specific exceptions for feature-limited workflows and simple API relays connected to independently run external model providers. An internal-use exception permits in-house enterprise deployment only if the underlying weights, derivatives and generated outputs are never supplied to outside third parties.
“The parameter total excludes the model’s vision encoder, adapter and DFly2 drafting component.”
Compute demands and the lighter Flash alternative
The public weight distribution follows the introduction of Tencent’s hosted Hy4 service on November 4 and the November 16 launch of Hy4 Flash, a condensed model intended to reduce enterprise running costs.
Hy4 Flash operates with 224 billion total parameters, of which 64 billion are dedicated Engram memory parameters based on an architecture originally proposed by DeepSeek. Flash activates 12 billion parameters per token and runs against a 1-million-token context window.
Tencent Cloud lists Hy4 Flash at $0.07 per million input tokens and $0.281 per million output tokens, alongside a cached input rate of $0.003 per million tokens. On domestic currency rate cards, the service costs 0.5 yuan per million input tokens and 2 yuan per million output tokens, with cache hits priced at 0.02 yuan.
Tencent stated that the combination of hybrid attention, a LatentMoE design and Engram parameters reduces the key-value cache footprint of Flash to one-sixth of flagship Hy4 and 1/24 of the older Hy3 generation. The provider reported decoding throughput exceeding 150 tokens per second when pairing Flash with megakernels and 4-bit quantization.
Benchmark performance and enterprise safety reviews
Task economics separate the two tiers. Tencent reported that Hy4 Flash completes standard tasks on its WorkIndex test suite at one-twentieth the operating cost of the 770-billion-parameter flagship and one-eighth the cost of rival DeepSeek-V4.1-Flash.
Testing across 259 workplace tasks by 85 WorkBuddy users rated Flash output close to Anthropic’s Claude Haiku 5.5. On technical benchmarks, Flash scored 19.2 on Terminal-Bench 4.0, while the flagship Hy4 model reached 44.4.
Alongside the flagship weight files, Tencent linked an official System Card detailing risk mitigations. The documentation highlights internal auditing by cybersecurity specialists and the removal of offensive computer exploit material from training datasets.
Those self-reported assessments leave businesses responsible for validating task reliability, hardware scaling and network safeguards before wiring the 770-billion-parameter architecture into operational enterprise stacks.
Questions & Answers
Q.What commercial restrictions apply to companies using the Hy4 model?
What commercial restrictions apply to companies using the Hy4 model?
Providers offering inference or fine-tuning services must obtain written permission from Tencent if their annual turnover exceeds $50 million, or if their product reaches 100 million monthly active users. In-house enterprise deployment is permitted only if the weights and generated outputs are never shared externally.
Q.What technical infrastructure is required to run the flagship Hy4 checkpoints locally?
What technical infrastructure is required to run the flagship Hy4 checkpoints locally?
Deploying the 770-billion-parameter flagship on private infrastructure requires substantial compute resources. Tencent’s published deployment documentation specifies vLLM and SGLang recipes that demand eight-way tensor parallelism to manage the mixture-of-experts model, which activates 49 billion parameters per token.
Q.How does the condensed Hy4 Flash model compare in benchmark performance and running costs?
How does the condensed Hy4 Flash model compare in benchmark performance and running costs?
Hy4 Flash completes standard workplace tasks at one-twentieth the operating cost of the flagship model. While the flagship scored 44.4 on Terminal-Bench 4.0 compared to 19.2 for Flash, workplace testers rated Flash output close to Anthropic’s Claude Haiku 5.5 across 259 tasks.
Q.What are the token pricing rates for the hosted Hy4 Flash service?
What are the token pricing rates for the hosted Hy4 Flash service?
Tencent Cloud lists Hy4 Flash at $0.07 per million input tokens and $0.281 per million output tokens, alongside a cached input rate of $0.003. Domestic rate cards list standard usage at 0.5 yuan for inputs, 2 yuan for outputs, and 0.02 yuan for cache hits.
Reader pulse
Will enterprises accept Tencent's commercial licensing restrictions?
21,214 votes so far