Overview of Tencent Hunyuan Large and 3D Generation Model Open‑Source Release

On November 5, Tencent announced the open‑source release of two core models: the Mixture‑of‑Experts (MoE) model "Hunyuan Large" and a 3D generation model. Both models can be downloaded from GitHub and Hugging Face, and are also accessible via the Tencent Cloud TI platform and the HAI high‑performance application service.

Hunyuan Large has a total parameter count of 389 B, 52 B active parameters, and a context length of up to 256 K tokens, making it the largest open‑source MoE model to date. Benchmark results show it leading in multi‑disciplinary evaluation suites such as CMMLU, MMLU, CEval, and MATH, outperforming leading open‑source models like Llama 3 and Mixtral.

The model introduces several MoE innovations:

Random‑compensation routing that redistributes tokens dropped by overloaded experts to under‑utilized experts, improving training stability.

A shared expert that captures common knowledge for all tokens, plus 16 routed experts that specialize based on activation scores.

High‑quality synthetic data pipelines and multi‑stage pre‑training with annealing to enhance long‑text capabilities.

KV‑cache compression using Grouped‑Query Attention (GQA) and Cross‑Layer Attention (CLA), reducing cache size to 5 % of a standard MHA and boosting inference speed.

Post‑training optimizations include a two‑stage reinforcement learning pipeline (offline DPO followed by online RL) and expert‑specific learning‑rate schedules, which together raise performance on mathematics, logic reasoning, and code generation.

The training framework AngelPTM implements expert‑compute/communication overlap, MoE operator fusion, and low‑precision training, delivering 2.6× the performance of DeepSpeed. For inference, the AngelHCF‑vLLM framework builds on vLLM with NF4 and FP8 quantization and parallel decoding, saving over 50 % of VRAM and achieving more than double the throughput of BF16‑based solutions.

Hunyuan 3D Generation Model is the first open‑source model that simultaneously supports text‑to‑3D and image‑to‑3D generation. Evaluated on the GSO and OmniObject3D datasets, it surpasses state‑of‑the‑art open‑source models in geometry detail, texture fidelity, consistency, realism, and instruction adherence. The model is already applied internally for UGC 3D creation, product material synthesis, and game asset generation.

Access links:

GitHub (model toolkit): https://github.com/Tencent/Hunyuan-Large

Hugging Face model card & upload: https://huggingface.co/tencent/Hunyuan-Large/tree/main

Demo on Hugging Face: https://huggingface.co/spaces/tencent/Hunyuan-Large

3D model GitHub: https://github.com/Tencent/Hunyuan3D-1

3D model demo: https://huggingface.co/spaces/tencent/Hunyuan3D-1

The release marks a major step in Tencent's open‑source strategy for large models, with future plans to open more models validated in real‑world Tencent applications.