Tencent

Compact single-codebook audio tokenizer unifying general audio perception and linguistic alignment for downstream Audio-LLMs. WhisperVQ encoder + flow…

"Exploring Autonomous Agentic Data Engineering for Model Specialization." Formalizes a new task — letting an LLM agent autonomously execute the…

First flagship model built under Chief AI Scientist Yao Shunyu (former OpenAI researcher, core contributor to Operator and Deep Research). 295B total…

Scaling flow matching models for text-to-motion generation. Enables high-quality human motion synthesis from text descriptions.

Next-generation foundation model with MoE architecture (406B total, 32B active) and 256K context window. Available in Think (deep reasoning, code gene…

Commercial-grade, lightweight (1B) end-to-end OCR expert VLM. Connects a 0.4B native-resolution ViT to a 0.5B Hunyuan LLM. Won first place in ICDAR 20…

Leading lightweight video generation model optimized for consumer GPUs.

All-in-one 3D reconstruction model that processes video to 3D in a single forward pass.

Native multimodal autoregressive model for image generation, built on the Hunyuan-A13B backbone (80B+ total, 13B active). Largest open-source image ge…

Open-source translation model series supporting 33 languages. Hunyuan-MT-7B won first place in 30 of 31 categories at WMT25. Hy-MT2 (May 21, 2026) ext…

Multimodal diffusion model with representation alignment for high-fidelity Foley audio generation synced to video content.

Interactive game video generation framework. Trained on 1M+ gameplay recordings across 100+ AAA games. Unifies keyboard/mouse inputs into a shared cam…

Research on unlocking flow-based GRPO (Group Relative Policy Optimization) efficiency for reasoning models.

Open-source world generation model series. World 1.0 integrates panoramic visual generation and 3D reconstruction from text/image inputs. HY-World 1.5…

High-efficiency open-source MoE model (80B total / 13B active) with dual-mode reasoning and 256K context. Frequently used in benchmarks and serves as…

Portrait animation and multi-character dialogue generation built on the HunyuanVideo framework. Enables audio-driven animated portraits.

Multimodal-driven architecture for customized video generation. Enables identity-preserving, style-consistent, and subject-driven video creation from…

Tencent's deep reasoning model, built on Hunyuan-TurboS with additional reinforcement learning for extended chain-of-thought reasoning.

Single-transformer baseline for multi-modal understanding, simplifying vision-language model architecture. Published at ICML 2025.

"Fast-thinking" model using a Hybrid-Mamba-Transformer architecture for near-instant replies with complex reasoning. 56B activated / 560B total hybrid…

Multimodal understanding model built on MoE architecture (389B total, 52B active). Handles images, videos, and 3D content. Ranked first among Chinese…

Scaling diffusion models for high-resolution textured 3D asset generation. 2.1 added PBR material synthesis for cinema-grade light interaction. 2.5 ac…

Diffusion-powered video tokenizer for comprehension and generation, unifying video understanding and synthesis. Published at CVPR 2025.

Latent motion tokens as a bridging language for learning robot manipulation from videos. Accepted at ICCV 2025 as Oral.

13B parameter systematic framework for high-fidelity video generation. Largest open-source video model at launch, with 3M+ HuggingFace downloads.

Scalable image tokenization with index backpropagation quantization. Accepted at ICCV 2025.

Open-source MoE model with 52B activated parameters out of 389B total. 256K context. Pre-trained on 7T tokens.

Unified framework for text-to-3D and image-to-3D generation.

Open-source replication and improvement of Google's MAGVIT-v2 visual tokenizer, achieving state-of-the-art image reconstruction quality.

Powerful open-source Diffusion Transformer (1.5B) with fine-grained Chinese understanding. Evolved through v1.1 and v1.2.

Comprehensive benchmark for evaluating multimodal LLMs on code generation from scientific plots. Published at NAACL Findings 2025.

Multimodal foundation model with unified multi-granularity comprehension and generation. Extends the SEED family to handle both fine-grained visual un…

Real-time open-vocabulary object detection, now a standard in computer vision. Accepted at CVPR 2024.

Progressive LLaMA with block expansion, enabling post-pretraining domain knowledge injection without catastrophic forgetting. Published at ACL 2024.

Large-kernel CNN architecture providing a state-of-the-art backbone for vision tasks. Published at CVPR 2024.

Toolbox for high-quality video diffusion models, including the widely used FreeNoise.

Multimodal foundation model enabling LLMs to "see" and "draw" via a discrete visual tokenizer. SEED-LLaMA pioneered multimodal in-context learning wit…

Debut flagship with hundreds of billions of parameters, initially used for internal products like Tencent Docs and Search.

Towards omni-modal representations by extending ViT to additional modalities (3D, audio, etc.) via lightweight lens modules. Published at CVPR 2024.

Benchmark for evaluating multimodal large language models across multiple dimensions. SEED-Bench-2 expanded to 24K multiple-choice questions covering…

Teaching large language models to use tools via self-instruction. Published at NeurIPS 2023.

Decentralized low-rank adaptation for multi-concept customization of diffusion models. Published at NeurIPS 2023.

One-shot tuning of image diffusion models for text-to-video generation. Influential early work on adapting image diffusion for video. Published at ICC…

Protein structure prediction model rivaling AlphaFold in accuracy, released for the life sciences community.

Quantum chemistry dataset for benchmarking AI models in molecular property prediction.

Tencent's first major AI release. Distributed machine learning platform based on a Parameter Server architecture. Later donated to the Linux Foundatio…