Hunyuan

Also known as: Tencent Hunyuan

On this page 7

Hunyuan, also called Tencent Hunyuan, is the family of foundation models developed by Tencent, covering large language models, reasoning models, image, video and 3D generation. Among AI models & assistants, it stands out as one of the largest Chinese model programs that combines closed cloud services with a steady stream of open-weight releases.

Hunyuan powers features across WeChat, Tencent Docs, Tencent Meeting and the Yuanbao assistant app, and it is sold to businesses through Tencent Cloud APIs. For developers outside Tencent, the name usually refers to specific downloadable checkpoints such as Hunyuan-Large, HunyuanVideo or Hunyuan3D, each with its own license terms.

What the Hunyuan model family includes

Hunyuan is not a single model but a brand that groups several model lines built by Tencent research teams since the first public announcement in September 2023. The lines share infrastructure, training data pipelines and naming, yet they target different modalities and are released on different schedules.

The following table lists the main lines that appear in public documentation and model repositories. Version numbers change often, so the table describes lines rather than exact releases.

Model lineModalityTypical availabilityMain purpose
Hunyuan LLM (Turbo, TurboS)TextTencent Cloud API, YuanbaoGeneral chat, writing, coding, search answers
Hunyuan-T1Text with reasoningAPI and assistant appStep-by-step math, logic and analysis
Hunyuan-Large and A13BTextOpen weightsSelf-hosted mixture-of-experts language model
Hunyuan-DiT and HunyuanImageImageOpen weights and APIBilingual text-to-image generation
HunyuanVideoVideoOpen weights and APIText-to-video and image-to-video clips
Hunyuan3D3D assetsOpen weights and web toolMeshes and textures from text or images
Main Hunyuan model lines and their purpose

This breadth is the defining trait of Hunyuan: Tencent uses one brand to cover nearly every generative modality, which makes direct comparison with a single text model such as Claude or Llama only partly meaningful.

How Hunyuan language models are built

The Hunyuan language models are decoder-only Transformers trained on large multilingual corpora with a strong emphasis on Chinese, followed by supervised fine-tuning and reinforcement learning from human feedback. Several recent versions use a mixture-of-experts (MoE) design, in which a router activates only a few expert subnetworks for each token.

The open Hunyuan-Large model, released in late 2024, was reported at about 389 billion total parameters with roughly 52 billion active per token. The later Hunyuan-A13B reduced this to about 80 billion total and 13 billion active, a size that fits on far fewer accelerators while keeping a long context window of up to 256,000 tokens.

The reasoning model Hunyuan-T1 and the fast TurboS line were described by Tencent as using a hybrid Transformer and Mamba architecture. Mamba layers are state-space blocks that process long sequences with near-linear cost, which aims to cut latency on long prompts and long chains of reasoning.

The table summarizes the architectural choices that matter most when evaluating these models. Figures are the publicly reported values and should be checked against the model card of the exact release.

FeatureWhat it meansPractical effect
Mixture of expertsOnly some expert layers run per tokenLower inference cost than a dense model of equal size
Hybrid Mamba layersState-space blocks mixed with attentionFaster decoding on long inputs
Long contextPrompts of 128,000 to 256,000 tokensWhole documents or code bases in one request
Fast and slow thinkingOptional extended reasoning modeTrade answer speed for accuracy on hard tasks
Architectural features of Hunyuan language models

Hunyuan generative media models

Hunyuan media models apply the diffusion transformer approach to images, video and 3D content. Hunyuan-DiT, released in 2024, was one of the first open text-to-image models with native Chinese and English prompts, which matters for prompts containing Chinese cultural details or text.

HunyuanVideo, published with open weights at the end of 2024, has about 13 billion parameters and generates short clips with consistent motion and camera control. It quickly became a common base for community fine-tuning, much like FLUX for still images, because the weights and inference code were public.

Hunyuan3D uses a two-stage pipeline: one model generates the geometry as a mesh, and a second model paints high-resolution textures onto it. Game studios and 3D printing hobbyists use it to create draft assets in minutes instead of hours of manual modeling.

How Hunyuan compares with other model families

Hunyuan is best compared with other large Chinese model families and with Western families that mix open and closed releases. The table below places it among neighbors from the same field, focusing on strategy rather than benchmark scores, which change with every release.

Model familyDeveloperOpennessDistinctive focus
HunyuanTencentMixed open and closedBroad modality coverage, WeChat ecosystem
QwenAlibabaLargely open weightsMany sizes, strong multilingual coverage
DeepSeekDeepSeekOpen weightsEfficient training, reasoning models
ERNIEBaiduMostly closed, some openSearch integration, knowledge graphs
LlamaMetaOpen weights with licenseGeneral text models for self-hosting
Hunyuan compared with related model families

A practical difference is distribution. Hunyuan reaches users mainly through Tencent products with over a billion accounts, so many people use it without knowing the model name. In early 2025 Tencent even added DeepSeek models inside Yuanbao next to Hunyuan, which shows that the assistant and the model are separate layers that can be combined.

Licensing, access and practical limitations

Open Hunyuan checkpoints are released under the Tencent Hunyuan Community License, not under a standard open-source license such as Apache 2.0. Several versions of this license exclude use in the European Union, the United Kingdom and South Korea and require a separate agreement above a large monthly active user threshold.

Teams evaluating Hunyuan should check the following points before deployment, because they cause most adoption problems in practice.

IssueCauseRecommended response
Territorial license limitsLicense excludes some regionsReview the license text for each checkpoint
Content filteringModels follow Chinese regulatory rulesTest sensitive topics before production use
Hardware demandsLarge MoE models need much memoryPrefer smaller or quantized releases
Sparse English documentationSome material appears in Chinese firstRely on official model cards and code
Fast version changesFrequent new releases and namesPin exact model versions in deployments
Common issues when adopting Hunyuan models

Benchmark claims in launch materials are self-reported by Tencent and often compare against selected competitors. Independent evaluations show Hunyuan models as competitive but not uniformly leading, with strength in Chinese-language tasks and visual generation.

In everyday language

Hunyuan is the name Tencent, the company behind WeChat, gives to its artificial intelligence models. Some of them chat and write, some solve math problems, and others draw pictures, make short videos or build 3D objects.

Think of Hunyuan as a family of specialists under one surname. Most people meet it inside apps they already use, such as a chat assistant or a document editor. Developers can also download some family members for free, but they must read the license, because it does not allow use everywhere.

Seen in the real world

A small independent game studio in Shenzhen needs about 200 background props for a stylized adventure game, including lanterns, crates, market stalls and statues. Its two artists cannot model all of them within the eight-week schedule.

The team runs Hunyuan3D on a local workstation with a 24 GB graphics card. An artist sketches or photographs a reference, generates a textured mesh in a few minutes, and then cleans the topology and adjusts colors in standard 3D software. Hero characters remain fully hand-made.

The result is a prop library finished in about three weeks instead of the planned seven. Because the studio sells the game only in Asian markets, the lead developer confirms that the community license terms fit their distribution before shipping.