News · 2026-11-18
Tencent’s Hy4 Flash puts a large checkpoint behind a conditional license
Tencent has published Hy4 Flash weights and a hosted catalog offering, giving developers access to a large multimodal model with an important hardware and licensing catch. The full-precision instruct repository’s listed files total about 440.5 GB, and Tencent’s license adds conditions for certain large commercial uses.
Key facts
- The Tencent model card describes 220 billion total parameters and 11.3 billion active text parameters per token.
- The repository file listing totals 440,533,677,833 bytes.
- Tencent documents eight-way serving; the linked recipe describes an eight-H200 configuration.
- The Hy License is based on Apache 2.0 but adds terms and is not standard Apache 2.0.
The model accepts text and images and generates text. Tencent’s card describes a mixture of experts: only selected components participate in processing each token, even though the full model contains many more parameters. Think of a large library that consults a small set of shelves for each question. The number of shelves visited does not determine the space needed to store the library.
That is why 11.3 billion active text parameters does not mean a small download. The listed full-precision instruct repository contains 148 numbered weight shards and other files totaling approximately 440.5 GB. Tencent also describes FP8 instruct and base variants, whose sizes should not be inferred from this total. The vLLM deployment recipe describes eight H200 GPUs and estimates roughly 529 GB aggregate GPU memory for full precision and 345 GB for FP8 in its stated configuration. Those are recipe estimates, not independently measured minimum requirements; host-memory offload can change the arrangement.
Tencent reports more than 150 generated tokens per second with its specified optimizations, including 4-bit quantization and a drafting technique. It also reports an internal comparison across 259 tasks involving 85 WorkBuddy users. These are company results, not an independently reproduced assessment. The precise hardware, workload and end-to-end experience matter when translating a speed claim into a production service.
The license says it is “not the Apache License 2.0.” It requires a separate written Tencent license for qualifying model-as-a-service businesses above $50 million in aggregate worldwide gross revenue in a consecutive 12-month period, or for use in products above 100 million monthly active users, as defined by its terms. It also describes an internal-use exception. Fine-tuned or quantized derivatives are included in the license’s definition of derivatives.
The model was on Hugging Face by November 17, according to repository creation metadata. An OpenRouter catalog record bears a November 16 timestamp, which does not prove exactly when an endpoint became usable. The community discussion contains both favorable reports and complaints about speed or translation, but those anecdotes do not establish general quality.
The practical choice is therefore broader than whether weights exist. A user must check the variant, storage, serving recipe and applicable license, then measure quality on their own tasks. Downloadable access provides meaningful control, but it does not make the model lightweight or its terms unrestricted.
The architecture includes several mechanisms that affect serving. Tencent describes sparse attention, sliding-window attention and a built-in draft component for speculative decoding. A draft model proposes a sequence of tokens and the main model checks the proposal, potentially reducing sequential work when enough proposals are accepted. The actual improvement depends on the workload and implementation; it cannot be inferred merely from the presence of the technique. The site’s speculative decoding lesson explains the general mechanism.
Tencent also describes Engram parameters and a mixture of routed experts. These explain why the model’s total parameter count, active parameters and memory needs are different quantities. A deployment may keep more of the model resident than it uses for an individual token, and additional memory is needed for runtime state. The recipe’s estimate is tied to its own context and concurrency settings, rather than a universal minimum.
The license thresholds should be read in the full text, including its definitions and internal-use exception. A company close to a threshold cannot conclude that every use is restricted solely from a headline, just as a user below one threshold should not assume that all other terms disappear. The key practical steps are to select the exact variant, inspect its files and serving guidance, and evaluate the license against the intended use. Hosted access avoids a local weight download, but introduces its own provider, privacy and service dependencies.
Key questions
How large is the Hy4 Flash download?
Can anyone use Hy4 Flash commercially without conditions?
Cite this
APA
Ground Truth. (2026, November 18). Tencent’s Hy4 Flash puts a large checkpoint behind a conditional license. Ground Truth. https://groundtruth.day/news/tencent-hy4-flash-open-weights-licensing.html
BibTeX
@misc{groundtruth:tencent-hy4-flash-open-weights-licensing,
title = {Tencent’s Hy4 Flash puts a large checkpoint behind a conditional license},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/tencent-hy4-flash-open-weights-licensing.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.