What happened
Tencent introduced Hy4 Preview, a new open-weight language model that accepts text input but no vision input. The company says it has 770 billion total parameters, 49 billion active parameters and a one-million-token context window. The source presents this as a substantial expansion over Hy3, which it says had 295 billion total parameters, 21 billion active parameters and a 256,000-token context window.
Tencent’s primary source introduces Hy4 Preview as a new open-weight large language model with text input only. It explicitly says the model does not accept vision input. The company describes the model as having 770 billion total parameters, of which 49 billion are active parameters. Those figures are claims made by Tencent in the source; the source does not include an independent audit or external testing of the architecture or parameter counts.
The source places Hy4 in the context of Tencent’s earlier Hy3 model. It says Hy3, released in July, had 295 billion total parameters, 21 billion active parameters and a 256,000-token context window. By comparison, Tencent presents Hy4 as a large increase in total and active parameters, along with a context window that is four times longer than the stated Hy3 window. The source does not provide additional information about the release date of Hy4 beyond describing it as being introduced today.
The release also exposes implementation details through a chat template associated with the model on Hugging Face. The template recognizes two stated reasoning settings: “high,” which the source says is the default, and “no_think,” which disables reasoning. The template includes error handling for other values. This indicates that the model’s serving configuration distinguishes between a reasoning-enabled mode and a mode intended to suppress that behavior, but the source does not explain how either mode is implemented internally or how their outputs compare.
The source says the author tested a prompt asking for an SVG of a pelican riding a bicycle through OpenRouter while using the default high-reasoning setting. It quotes a short portion of the model’s apparent reasoning trace, in which the system considers whether to add a helmet, sunglasses or water to the requested image concept. This is an anecdotal example of text generation and planning behavior, not a controlled evaluation. It does not establish the model’s reliability, artistic quality, reasoning ability or suitability for production use.
Hy4 is described as open weight, but the source does not specify the precise weight files, license, supported hardware, quantization options, training data, safety controls or distribution terms. It also does not say whether Tencent is offering a hosted API, what regions or users can access the model, or whether the preview is stable. Those omissions are important for anyone attempting to reproduce the reported experience or deploy the model.
Source details: hy.tencent.ai ↗
Why it matters
The release points to continued competition among large language-model developers over model scale, sparse activation and context length. If Tencent’s stated specifications are accurate, Hy4 could give developers access to a model with a very large working context while activating a smaller portion of its total parameters for each task. The practical value remains uncertain because the source provides no independent benchmark results, pricing, service guarantees or detailed license terms.
The most consequential stated feature is the combination of open weights and a one-million-token context window. A longer context can allow a model to process larger collections of text in a single interaction, such as extensive code repositories, lengthy technical materials or multiple related records. That potential is conditional: the source does not report long-context accuracy, retrieval quality, degradation over distance or the cost of processing such inputs.
The parameter figures also illustrate the tradeoff between model scale and per-request computation. Tencent says Hy4 contains 770 billion total parameters but activates 49 billion at a time. In principle, sparse activation can let a model use a very large parameter pool without computing across all parameters for every token. The source does not provide inference speed, memory requirements, hardware guidance or measured cost, so it is not possible to determine from this announcement how accessible the model will be outside well-resourced deployments.
The open-weight framing could matter for researchers and developers who need more control than a closed hosted service provides. Access to weights may support local testing, adaptation and independent evaluation, depending on the license and technical requirements. But “open weight” does not by itself establish that the training data, training code, evaluation process or deployment stack are open. The source gives no information about those other layers of openness.
The reasoning controls may also be practically relevant. A default high-reasoning mode could affect response latency, token usage and output style, while a no-think mode could be useful when users want shorter or more direct responses. However, the source offers no measurements showing whether the setting improves accuracy, reduces cost or changes failure patterns. The presence of a configuration option should therefore not be treated as evidence of superior reasoning performance.
For the public, the announcement is mainly significant as a potential expansion of access to a very large language model from Tencent. It may broaden the range of systems available for comparison and experimentation, particularly in China-linked and open-weight model ecosystems. Yet the practical impact cannot be assessed fully from the company’s specifications alone. No benchmark scores, safety findings, user numbers, commercial deployments or independent reactions appear in the source.
What to watch next
The key next questions are whether Hy4’s weights are broadly downloadable, what license governs commercial and research use, and how the model performs against comparable systems on coding, reasoning, multilingual work and long-context retrieval. Developers should also watch actual memory, latency and serving requirements, since total parameter count can materially affect deployment even when only a subset is active. The source does not establish those operational details.
The first verification priority is the model’s actual availability. The source refers to open weights and points to a Hugging Face chat template, but it does not identify a complete download package, a formal license or access requirements. Developers should confirm which files are available, whether the weights are complete or restricted, and whether commercial use is permitted.
Independent evaluations will be needed to test the claims that matter most in practice. Useful checks would include coding and reasoning accuracy, multilingual performance, long-context retrieval, hallucination rates, instruction following and the difference between high and no-think modes. The source contains only one informal SVG-related interaction, which cannot answer those questions.
Deployment requirements are another major unknown. A 770-billion-parameter model may impose substantial memory and infrastructure demands even if 49 billion parameters are active for an individual operation. Until Tencent or independent users publish hardware, quantization, throughput, latency and cost information, the model’s accessibility for ordinary developers and smaller organizations will remain unclear.
The model’s safety and governance documentation also warrant attention. The source does not describe refusal behavior, harmful-content safeguards, privacy practices, training-data provenance or methods for reporting problems. Those details will be important if Hy4 is used in public-facing services, software development or other settings where inaccurate or unsafe outputs can have consequences.
Finally, readers should distinguish the preview’s stated specifications from demonstrated capability. Tencent’s announcement establishes that the company is presenting Hy4 as an open-weight, text-only model with the listed parameter and context figures. It does not establish that the model matches those claims on independent tests, that it is broadly usable today, or that its larger scale translates into better results. Further documentation and external evaluation will determine whether the release is a practical step forward or primarily a specification milestone.