Tencent's Hy4 Weights Are on Hugging Face: 770B Total, 49B Active
The Hy4 instruct weights on Hugging Face, under tencent/Hy4, sit in a commit timestamped 2026-11-20T03:05:49Z. That revision contains 245 safetensors shards. GitHub's Tencent-Hunyuan/Hy4 added the release README, the license, the benchmark charts, and the finetune scripts at 2026-11-19T12:51:29Z, with smaller README edits through 2026-11-20T02:46:38Z. Tencent's research post is dated November 4 and already calls Hy4 open-weight. The specifications below are from the model card.
The numbers
- 770 billion total parameters, excluding the vision transformer, the adapter, and the DFly2 drafter. 49 billion parameters activated per token for text, plus a 1.23 billion parameter vision encoder. Model card.
- Context length of 1,048,576 tokens. 78 backbone layers: one dense layer and 77 mixture-of-experts layers. Each MoE layer has 256 routed experts and one shared expert. Each token uses the top 8 routed experts plus the shared expert. Model card.
- 245 safetensors shards in the November 20 Hugging Face revision, named model-00001-of-00245 through model-00245-of-00245. Hugging Face also labels the repo 774B. The card's 770 billion figure leaves out the vision encoder, the adapter, and DFly2. This article does not add those into a new total.
- API prices on the November 4 research post: $0.042 per million cached input tokens, $0.834 per million input tokens, and $2.501 per million output tokens. The post says that is the same price as Hy4 preview.
What is in the release
The card names three variants: Hy4, the instruct model; Hy4-FP8; and Hy4 base. It says the base model is pretrained and not instruction-tuned. A 5-layer DFly2 drafter of 3.3 billion parameters is built in for speculative decoding. The card recommends temperature 0.9 and top_p 1.0, and says reasoning effort defaults to high.
The license is the Tencent Hy License, version 1.0. The text says it is derived from Apache 2.0 and that it is not the Apache License. Section 10 requires a separate written license before a model-as-a-service business uses the weights if worldwide gross revenue exceeds $50 million in any consecutive 12 months, whether or not that revenue comes from the model, or if a product has more than 100 million monthly active users in any calendar month. Internal use is exempt when the weights, derivatives, and outputs are not made available to third parties.
Set next to two other releases already on this site: Alibaba's Qwen4-Max kept its flagship weights closed, and Meituan shipped LongCat-2.5 as open weights. Hy4's card puts the active-parameter count in the same neighborhood as LongCat's, on a different total.
Benchmarks Tencent prints
These scores are Tencent's. On the research post, 213 finance professionals scored 579 tasks. Hy4 averaged 3.04 out of 4. The same table lists 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3. The post says the panel was 213 finance professionals, including 153 from financial institutions, and that more than 70 percent of these experts have over five years of experience. It says an internal search evaluation cut the hallucination rate from 15.76 percent on Hy4 preview to 5.45 percent. Against Hy4 preview, it reports a 19.1 percent gain on GameDev and a 9.8 percent gain on GameLogic. On Hy-BioExperimentBench it reports 30.4 out of 100 and says that sits below Tencent's own high-risk threshold.
The model card's base-model table gives Hy4 Base a 5-shot MMLU of 91.6 and an MBPP-plus score of 74.0, against 79.1 for DeepSeek-V4-Pro-Base on that code row. The card says the base model saw a large amount of synthetic instruction data in mid-training, and that this can distort few-shot scores when the model misses the expected answer format. The system card PDF linked from the model card did not load for this desk, so the safety figures here are from the research post, not from that PDF.
What is not disclosed?
The model card text retrieved here does not give a training-token count. The commit timestamp is for the revision as a whole. It does not say when each of the 245 shards was first uploaded. Tencent's transparency center lists Hy4 under November 4, 2026 and Hy4 flash under November 16, 2026. The research post, still dated November 4, says the weights are open and also says a Hy4 flash release is still ahead. The file dates a developer can check are November 19 on GitHub and November 20 on Hugging Face.
Quick answers
- What happened with Tencent?
- The Hy4 instruct weights on Hugging Face, under tencent/Hy4, sit in a commit timestamped 2026-11-20T03:05:49Z. That revision contains 245 safetensors shards. GitHub's Tencent-Hunyuan/Hy4 added the release README, the license, the benchmark charts, and the finetune scripts at 2026-11-19T12:51:29Z, with smaller README edits through 2026-11-20T02:46:38Z. Tencent's research post is dated November 4 and already calls Hy4 open-weight.
- What are the key numbers?
- 770 billion total parameters, excluding the vision transformer, the adapter, and the DFly2 drafter. 49 billion parameters activated per token for text, plus a 1.23 billion parameter vision encoder. Model card.
- What is not disclosed?
- The model card text retrieved here does not give a training-token count. The commit timestamp is for the revision as a whole. It does not say when each of the 245 shards was first uploaded. Tencent's transparency center lists Hy4 under November 4, 2026 and Hy4 flash under November 16, 2026. The research post, still dated November 4, says the weights are open and also says a Hy4 flash release is still ahead.
This article was ultrathought.