Tencent has made the weights of its flagship model Tencent Hy4 available for download. The multimodal mixture-of-experts model has 770 billion parameters, 49 billion of which are active per token, and handles a context of up to one million tokens. Three variants are now on Hugging Face, ModelScope and GitCode: the instruct model, an FP8 version and the pre-trained base model. The catch is the licence. Unlike the preview release, Hy4 no longer ships under Apache 2.0 but under Tencent’s own licence with extra terms for large providers.
Tencent Hy4: What the open weights contain
The key figures come from the model card on Hugging Face. The language backbone has 78 layers. The first uses a conventional dense feed-forward network, the remaining 77 are MoE layers with 256 routed experts and one shared expert each. Every token activates the eight best-matching experts plus the shared one. On top of that comes a 1.23-billion-parameter vision encoder that turns images at dynamic resolutions and video frames into tokens.
For attention, Tencent says it took inspiration from DeepSeek and GLM. Hy4 uses “Gated DeepSeek Sparse Attention” with multi-head latent attention. An indexer picks the 2,048 most relevant tokens from the context for each query instead of looking at all of them, which is meant to keep compute in check for very long inputs. “IndexCache” additionally reuses these selection indices across layers.
- Total parameters: 770 billion (excluding vision encoder and drafter)
- Active per token: 49 billion for text plus 1.23 billion for vision
- Context window: 1,048,576 tokens
- Experts: 256 per MoE layer, 8 active, plus 1 shared expert
- Drafter: DFly2 with 3.3 billion parameters for speculative decoding
The built-in DFly2 drafter matters for deployment. It proposes tokens that the large model then verifies, which speeds up generation. According to Tencent, it shares the KV cache and sparse indices with the main model. Each draft token therefore attends to at most around 2,000 keys regardless of context length, and no extra memory is needed for a separate cache.
Three checkpoints, 1.5 terabytes of data
Running Hy4 yourself takes serious storage. According to the Hub, the Hugging Face repositories for the instruct and base models each take up around 1.5 terabytes, the FP8 variant about 0.79 terabytes. For production serving, Tencent recommends vLLM or SGLang, both of which have dedicated prebuilt container images. The example configurations in the model card launch the FP8 version with tensor parallelism across eight GPUs. This is not a model for a desktop machine.
For testing, Tencent also offers a GGUF version via its AngelSlim compression toolkit, explicitly “for end-to-end testing purposes”, along with a fine-tuning pipeline. The reasoning mode is configurable: by default Hy4 uses an extended chain of thought (“high”), while “no_think” returns direct answers.
The model itself is not brand new. Tencent introduced Hy4 on November 4 and offered it as an API through Tencent Cloud and OpenRouter. Tencent lists prices of $0.834 per million input tokens, $0.042 for cached input and $2.501 per million output tokens. What is new is that the weights are now public. Its smaller sibling, Hy4 flash, with 220 billion parameters and 12 billion active, was released as an open-weight model two days earlier.
Benchmarks: strong on knowledge, weaker on code
For the base model, Tencent publishes a table of few-shot results comparing Hy4 Base with Hy4 flash Base, the older Hy3 preview and two DeepSeek base models. Hy4 leads on many knowledge tests, for example MMLU at 91.6 versus 91.1 for DeepSeek-V4-Pro-Base, or the SimpleQA factuality test at 57.2 versus 52.1. On the MBPP-plus coding benchmark, however, Hy4 Base scores only 74.0 while DeepSeek-V4-Pro-Base reaches 79.1. All figures were measured by Tencent itself; independent reproductions are still pending.
Tencent flags one caveat itself: because the base model saw a lot of synthetic instruction data during training, it occasionally ignores the expected answer format in few-shot evaluations, which may distort the scores. The model card is unusually candid elsewhere too. Under “Known Limitations”, Tencent concedes that a notable gap remains between Hy4 and the top proprietary models, particularly on research-level mathematics, long-horizon coding and advanced visual question answering.
For the instruct model, Tencent mainly cites use cases from inside the group. In an evaluation with 213 finance professionals across 579 real-world financial tasks, Hy4 averaged 3.04 out of 4 points, ahead of DeepSeek-V4.1-Flash (2.97), Kimi K3.1 Pro and GLM-5.3 (2.68 each). According to internal measurements, the hallucination rate on research tasks fell from 15.76 percent for Hy4 preview to 5.45 percent. In game development, for which Tencent curated training data with its gaming division IEG, the model improved by 19.1 percent on GameDev and 9.8 percent on GameLogic.
The licence: open, but not Apache
The most important change for companies sits in the licence file. Hy4 preview was released under Apache 2.0. Hy4 now uses the “Tencent Hy License, Version 1.0”. It is derived from Apache 2.0 but adds commercial terms for so-called model-as-a-service providers in Section 10, meaning companies that give third parties API access to inference or fine-tuning.
Such a provider needs a separate written licence from Tencent if its worldwide gross revenue exceeds $50 million in any 12-month period, whether or not that revenue comes from Hy4. The same applies before using the model in a product with more than 100 million monthly active users. Purely internal use is exempt, as long as neither the model, any derivative nor their outputs reach third parties. End-user products that confine model capabilities to specific features, without general-purpose model access, fall outside the MaaS definition.
For the community, one point stands out: fine-tuned, quantised, distilled, merged or pruned versions explicitly count as derivative works. Anyone publishing a Hy4 derivative passes on the same terms. For start-ups, researchers and most companies, little changes. For large cloud and API providers, however, Hy4 is no longer a freely exploitable model.
Context: the race of open giant models
Hy4 adds to a growing number of very large Chinese models with openly available weights. Only days ago, Alibaba presented a new flagship with Qwen4-Max, while Baidu is betting on an especially lean model with ERNIE 5.5 Flash. Tencent positions Hy4 between these worlds: large enough to keep up with the strongest open models on knowledge tests, with Hy4 flash as a cheaper option for everyday tasks.
For developers, Hy4 is mainly interesting where data must not leave their own data centre and enough GPU capacity is available. Teams with smaller budgets will more likely end up with Hy4 flash or other compact open-weight models such as Cloudflare’s Clairic. Before going to production, a close look at Section 10 of the licence is essential in any case. The benchmarks are a promise; the licence is a contract.



