Tencent releases downloadable weights for its 770-billion-parameter flagship Hy4
Edited and reviewed by AI Ping Editorial Team
3-line summary
- Hy4, its FP8 version and pretrained base weights are available for download.
- Large MaaS businesses face separate written-permission conditions tied to revenue or users.
- Serving and fine-tuning documentation targets servers with multiple GPUs.
Tencent has made its flagship Hy4 model weights available for download. Its official repository, checked on November 20, lists instruction-tuned Hy4, the smaller-file Hy4-FP8 version and a pretrained base checkpoint. Developers can now host the model themselves alongside the existing route of calling a cloud API.
The language backbone has 770 billion parameters, activating 49 billion per token through a mixture-of-experts architecture. It supports text, image and video understanding with a 1,048,576-token context. Hy4 base precedes instruction tuning; the chat examples apply to Hy4 and FP8. These are separate checkpoints from the smaller Hy4 flash. A 1.23-billion-parameter vision encoder and 3.3-billion-parameter DFly2 draft model sit alongside the backbone. The drafter proposes subsequent output while sharing the main model’s context cache. Tencent acknowledges gaps against leading proprietary models in research mathematics, long-horizon coding and advanced visual questions.
Additional terms for large model API businesses
Tencent Hy License 1.0 adds commercial conditions to Apache 2.0. MaaS inference or fine-tuning operators need separate written permission above $50 million worldwide gross revenue in any consecutive twelve months, or for products exceeding 100 million monthly active users. Revenue includes non-Hy4 income.
Feature-specific products without general model access are excluded from MaaS. Internal use is excepted when neither weights, derivatives nor outputs reach third parties.
Self-hosting requires substantial memory. The vLLM recipe specifies a single node with eight NVIDIA GPUs, around 944GB of aggregate GPU memory for FP8 or 1,859GB for BF16. Its checkpoint-storage figures are 786GB and 1,549GB respectively. These are requirements stated for that configuration, rather than measurements from an AI Ping deployment.
The recipe requires a Hy4-specific image or nightly builds from November 20 onward: stable vLLM 0.34.0 does not register the architecture. Its default evaluated window is 131,072 tokens with 32 concurrent sequences. The advertised model context therefore exceeds this serving configuration’s validated scope; a full one-million-token request has not been validated for the recipe.
Tencent’s fine-tuning guide lists tested minimums of eight machines and 64 GPUs for LoRA, or sixteen machines and 128 GPUs for full fine-tuning. Each GPU needs at least 96GB of memory and each machine at least 2TB of CPU memory. Longer sequences, larger batches, higher image resolution and more video frames raise requirements. The release includes mixed text, image and video training formats, with paths for DeepSpeed, LLaMA-Factory and ms-swift. Omitted reasoning settings default to fast responses during training, whereas the inference chat template defaults to deep thinking.
Cloud access remains available. Together AI announced serverless Hy4 on November 19, at $0.83 per million input tokens, $0.042 for cached input and $2.50 for output. Those API prices have a different basis from acquiring hardware to run downloaded weights. Its service supports image input, tool calling and JSON output.
Together published two demonstrations: reading a chart’s title and three values into JSON before rebuilding it in HTML, and generating a browser game from text. It says the unchanged game completed a round, reached game over and restarted. These selected examples establish neither a general task success rate nor performance in a self-hosted deployment.
Sources
- Hy4 official repository and model specifications — Tencent
- Tencent Hy License, Version 1.0 — Tencent
- Hy4 serving recipe — vLLM
- Hy4 Model Fine-tuning guide — Tencent
- Tencent Hy4 is live on Together AI — Together AI
- Hy4 public checkpoint metadata — Tencent / Hugging Face