What shipped
The Hy4 that launched on November 4 was an API product. As ALFW reported at the time, the Hugging Face repository was not yet public, and the open-source claim rested on a promise. That promise is now kept. All three checkpoints are public, and the model card carries the full specification: a 78-layer backbone with one dense layer and 77 MoE layers, sparse attention that selects the 2,048 most relevant positions per token so the million-token window stays affordable, and a built-in 3.3B-parameter DFly2 drafter for speculative decoding that reuses the main model's KV cache.
The practical barrier is size. At roughly 1.55 TB in BF16, or 0.79 TB in FP8 — file sizes per the release coverage — this is a data center download, and the coverage reports Tencent's vLLM recipe spreading the FP8 checkpoint across eight GPUs with tensor parallelism. The difference the release makes is that running it yourself is now possible in principle, and researchers can inspect what the API hides.
The license is the story
Hy4 preview, released in August, shipped under Apache 2.0, per the earlier coverage. Hy4 ships under the Tencent Hy License 1.0. The license is derived from Apache 2.0, and its text says plainly that it is not Apache 2.0. The added Section 10 covers model-as-a-service: giving third parties remote access to inference or fine-tuning, whether or not they pay. A business doing that needs a separate written license from Tencent if its aggregate worldwide gross revenue exceeds US$50 million in any 12-month period, or if the product using the model exceeds 100 million monthly active users in any calendar month. Internal use is exempt, provided the model, its derivatives and their outputs do not reach third parties. Fine-tuned, quantized, distilled, merged and pruned versions all count as derivatives, and Section 10 continues to apply to every recipient of a derivative work.
The Flash sibling, released November 16, uses the same license, so both Hy4 releases now carry Tencent's standard terms for the line rather than a one-off.
The numbers are Tencent's
The model card's base-model table is strong where Tencent chose to compete. Hy4 Base leads the listed comparison on MMLU (91.6), MMLU-Pro (73.9), SuperGPQA (62.7) and SimpleQA (57.2, against 26.4 for Hy3 preview in the same table). Its multimodal scores, MMMU-Pro 59.6 and CVBench 86.7 among them, are reported only against other Hy models, so they rank the family, not the field. On MBPP-plus, Hy4 Base trails the strongest listed rival: 74.0 against DeepSeek-V4-Pro-Base's 79.1 in Tencent's own table. Tencent also concedes a remaining gap to closed models on research-level mathematics, long-horizon coding and advanced visual question answering, and warns that synthetic instruction data in mid-training can make the base model miss expected answer formats in few-shot tests.
None of this has been independently reproduced. That was true of the November 4 launch coverage too, and it stays true until a third party runs the checkpoints, which, as of today, they finally can.
What to do with it
If you evaluate open-weight flagships, Hy4 now belongs on the list, and the evaluation can now be run honestly: download the FP8 checkpoint and test it on your own tasks instead of quoting the card, if you have the hardware. If you build a product on third-party API access to Chinese open models, put the license in the diligence file now, while you are small enough that Section 10 does not bind. The triggers are the operator's revenue or the product's user count, not the volume of inference served.
Sources
- [1] Tencent Hy4 model card, Hugging Face (specifications, benchmark table, checkpoint list, license badge)Read source
- [2] Tencent Hy License 1.0, the LICENSE file in the Hy4 repository on Hugging Face (license text, Section 10; label only, no separate URL recorded)
- [3] Sulat, "Tencent releases Hy4 open weights: 770B MoE, custom licence," November 20, 2026Read source
- [4] AlphaSignal, "Tencent's Hy4 Opens a 770B Model With a 1M-Token Window," November 20, 2026Read source