The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

Tencent Releases Hy4 AI Model Weights for Download

By The Daily Byte ·21 Nov 2026 ·3 min read
Tencent headquarters twin towers in Shenzhen

Tencent has published the full weights of its flagship Hy4 artificial intelligence model, allowing developers to download and deploy the system themselves. The release, published on Friday 20 November on Hugging Face, ModelScope and GitCode, covers the instruct model, a lower-precision FP8 build and the pre-trained base model. It comes a fortnight after Hy4 first reached Tencent's own application programming interface on 4 November, when the weights were promised but not yet available.

Hy4 is a mixture-of-experts model with 770 billion total parameters, of which 49 billion are active for each token. It is built from 78 layers with 256 routed experts, eight of which fire per token, and it supports text, image and video understanding with a context window of 1,048,576 tokens. Tencent says the model improves on the earlier Hy4 preview in instruction following, long-context understanding, extended tasks and research, and it can operate software through graphical interfaces and command-line tools, alongside front-end development, game development and office work.

What developers actually get

The download is not a small one. The full-precision checkpoint is about 1.55 terabytes, the FP8 build about 0.79 terabytes and the base model about 1.54 terabytes, plus a system card. The 770-billion figure covers the language backbone alone and excludes a 1.23-billion-parameter vision encoder and a 3.3-billion-parameter speculative-decoding component, known as DFly2, which reuses the model's own memory cache. Tencent publishes serving recipes for the vLLM and SGLang frameworks, using tensor parallelism across eight graphics processors, and ships a prebuilt vLLM image for the FP8 build. In practice, an eight-GPU server of the H200 or B200 class is the realistic serving unit for the FP8 checkpoint; at full precision the job becomes a multi-node one.

The licence carries a catch

The release does not ship under the plain Apache 2.0 licence used for Tencent's earlier Hy models. It uses the Tencent Hy License, whose section 10 requires a separate written licence from any operator selling Hy4 inference as a service once its worldwide gross revenue passes 50 million US dollars in any twelve months, or its product passes 100 million monthly active users. Purely internal use is exempt, and below those thresholds the weights are free to download and run. One early report described the release as Apache 2.0, but the model card itself carries the Tencent Hy License, and that is the document developers must follow. The same clause shipped with the smaller Hy4 flash model, released on 16 November.

Benchmarks are Tencent's own

Every benchmark on the model card is Tencent's own evaluation, not an independent result. The base model scores 91.6 on MMLU and 73.9 on MMLU-Pro in Tencent's tables, ahead of the DeepSeek and in-house baselines it lists. The card is also candid about the gap that remains to the leading proprietary models on research-level mathematics and long-horizon coding. Until independent testing appears, the figures should be read as the maker's measurements.

Why this release matters

The publication closes the open question left by the 4 November launch, when Hy4 was available only as a hosted service through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy and Yuanbao, and through gateways including OpenRouter. The weights were promised for the coming days, with no date attached, and only the August preview weights were downloadable. Now developers can run the flagship themselves, fine-tune it and inspect it, subject to the licence threshold. The practical barrier has shifted from access to hardware: at roughly three-quarters of a terabyte even in its compressed form, Hy4 is a data-centre model, not a laptop download.