Tencent has published the full weights of Hy4, its 770-billion-parameter flagship model, letting anyone download and run it. The release, posted on Friday, covers the instruct model, an eight-bit FP8 build and the pre-trained base, all free on Hugging Face.
What Tencent Released
The three checkpoints sit on Hugging Face, ModelScope and GitCode, alongside a system card with safety evaluations. The model itself is a mixture of experts: 770 billion parameters in total, with 49 billion active per token, across 78 layers and 256 routed experts.
Hy4 reads text, images and video, and handles a context window of about one million tokens. It also works the computer directly, through command-line and graphical interfaces, a capability Tencent first showed when the model launched as a paid service on 4 November.
A built-in drafter called DFly2 speeds up generation by reusing the model’s own memory, so long-context answers cost less to produce. Tencent ships official recipes for vLLM and SGLang, plus a compressed GGUF build, and says the FP8 checkpoint runs on eight graphics cards.
The Licence Has a Catch
The weights are not under Apache 2.0, despite some early reports saying otherwise. Tencent’s own model card states plainly that Hy4 ships under the Tencent Hy License, a custom licence derived from Apache 2.0 but with extra commercial terms attached.
The key term sits in Section 10: anyone selling Hy4 inference as a service needs Tencent’s separate written permission once revenue passes $50 million or the product tops 100 million monthly users. Purely internal use is exempt, and smaller operators can run it freely.
How It Measures Up
On Tencent’s benchmarks, the base model scores 91.6 on MMLU and 73.9 on MMLU-Pro, ahead of DeepSeek’s V4 Pro base on most tests. In a finance evaluation judged by 213 professionals, Hy4 ranked first among open-weight models with 3.04 out of 4, ahead of Kimi’s K3.1 Pro at 2.68.
Those figures are the maker’s own measurements, not independently reproduced results, and the card is candid about the gaps that remain. Hy4 still trails the best proprietary models on research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.
Our Read
This is the open-weight frontier getting heavier. A 770-billion-parameter model with vision, computer use and a million-token window is now downloadable, and that puts serious capability within reach of any team with a multi-GPU server. The catch is the licence, not the size.
Tencent is redrawing what open means: free to run, but not free to sell at scale without asking first. That is a different bargain from DeepSeek’s fully open releases, and it protects Tencent’s own cloud business while still winning the goodwill of developers who want the weights.
Why It Matters
For teams building their own agents, the release matters because self-hosting removes the meter: no per-token bill, no vendor lock-in, and full control over data. For the wider race, it keeps China’s open-weight labs shoulder to shoulder with the West’s best freely available models.
Background
Hy4 first appeared in preview in August under Apache 2.0, then launched in full on 4 November as a paid service, with the weights promised but not yet released. Tencent opened the smaller Hy4 Flash weights two days ago, and the flagship’s arrival now completes the family’s open release.
Source: Tencent’s Hy4 model card and TechNode’s report


0 Comments