Latest22 Sep 2026Image model
Hy Image3.5 preview
Tencent Hunyuan's newest image generation and editing model, released as a preview for professional image work. It handles text-to-image, image-to-image and multi-turn editing in conversation at up to 2K resolution. Tencent says its in-house designers rated it 30% better than Hy Image3.0 in blind tests.
- Text-to-image and image-to-image generation
- Up to 5 reference images per request
- Multi-turn conversational creation and editing
- Output up to 2K resolution
- 30% above Hy Image3.0 in Tencent internal designer blind test
AccessTencent apps including Yuanbao and ima; API through Tencent Cloud TokenHub. Not open weights.
PricePer image, not per token: CNY 0.15 per 2K image (Tencent Cloud); US$0.024 per image (Tencent Cloud international)
9 Sep 2026Audio/voice model
AuK
An open-source foundation model for generating and editing speech, released with a distilled AuK-Flash variant that runs in four steps. It covers text-to-speech, edits to content, pitch, speed, emotion and timbre, speech enhancement and source separation. Code and weights are under the MIT licence.
- 1.5B parameters (base model)
- Zero-shot and instruction-based text-to-speech
- Edits emotion, timbre, accent, whisper, nonverbal sounds
- Denoising, music and speaker separation
- MIT licence; English and Mandarin demos
AccessOpen weights and code on GitHub/Hugging Face (MIT licence)
28 Aug 2026Open-weights LLM
Hy4 preview
A preview of Tencent's next-generation model: a 770B-parameter mixture-of-experts system with a context window of about one million tokens. It is aimed at coding, agent and research work. Tencent released the weights under Apache 2.0 at the same time as API access.
- 770B total / 49B active parameters (MoE, 78 layers)
- 1M-token context
- Apache 2.0 open weights, incl. FP8
- SWE-Bench Pro 65.7, GPQA Diamond 92.3 (model card)
AccessOpen weights (Apache 2.0) on Hugging Face, ModelScope, GitCode, CNB; Tencent Cloud TokenHub API; OpenRouter; Yuanbao, ima, WorkBuddy, CodeBuddy
6 Jul 2026Open-weights LLM
Hy3
The final release of Tencent's rebuilt Hy3 model, following the April preview with more post-training compute and better data. It is a 295B mixture-of-experts model with a 256K context. Tencent open-sourced it under Apache 2.0 and built it into its Yuanbao assistant, with a strong focus on agent tasks.
- 295B total / 21B active parameters (+3.8B MTP layer)
- 256K-token context
- Apache 2.0; FP8 version available
- SWE-bench Verified 78%, GPQA Diamond 90.4 (model card)
AccessOpen weights (Apache 2.0) on Hugging Face, ModelScope, GitCode, CNB; Tencent Cloud API; Yuanbao app; OpenRouter
21 May 2026Open-weights LLM
Hy-MT2
Tencent's second-generation open translation models, in 1.8B, 7B and 30B-A3B (mixture-of-experts) sizes. They cover 33 languages and follow instructions given in multiple languages. Tencent says aggressive quantisation shrinks the 1.8B model to 440MB for on-device use, and released a new benchmark with them.
- Sizes: 1.8B, 7B, 30B-A3B MoE
- Translation across 33 languages
- 1.25-bit quantisation: 1.8B model in 440MB
- Released with the IFMTBench benchmark
AccessOpen weights on GitHub/Hugging Face (incl. GGUF builds)
23 Apr 2026Open-weights LLM
Hy3 preview
The first model from Tencent's rebuilt pre-training and reinforcement-learning stack, released as an open-weights preview. It is a 295B mixture-of-experts model with 21B active parameters and a 256K context, and it was already deployed in Tencent's Yuanbao chatbot at launch.
- 295B total / 21B active (+3.8B MTP layer)
- 256K-token context
- Tencent Hy Community licence
- Deployed in Yuanbao at launch
AccessOpen weights on Hugging Face (Tencent Hy Community licence); Yuanbao; OpenRouter
16 Apr 2026World model
HY-World 2.0
A multimodal world model that reconstructs, generates and simulates navigable 3D worlds from text, single or multiple images, or video. It outputs Gaussian splats, meshes and point clouds that work in Blender, Unity and Unreal. Tencent released the technical report first and the code and weights in stages.
- Inputs: text, single/multi-view images, video
- Outputs 3DGS, meshes, point clouds, depth, normals
- HY-Pano 2.0 (~80B), WorldStereo 2.0 (~17B), WorldMirror 2.0 (~1.2B)
- Exports for Blender, Unity and Unreal
AccessCode and weights on GitHub/Hugging Face (tencent-hy-world-2.0-community licence), released in stages
5 Dec 2025Frontier LLMReported
Tencent HY 2.0 (Hunyuan 2.0)
Tencent's proprietary flagship at the end of 2025: a 406B mixture-of-experts model released in Think and Instruct versions. Tencent reported gains in maths, science and coding reasoning from better pre-training data and reinforcement learning, and rolled it into its Yuanbao and ima apps and Tencent Cloud APIs.
- 406B total / 32B active parameters (MoE)
- 256K-token context
- Think (reasoning) and Instruct versions
- Available via Yuanbao, ima and Tencent Cloud API
AccessYuanbao and ima apps; Tencent Cloud API
20 Nov 2025Video model
HunyuanVideo-1.5
A lightweight open video generation model of 8.3B parameters for text-to-video and image-to-video. It produces 480p and 720p clips of about five seconds, with an optional super-resolution step to 1080p. Tencent later added step-distilled and cache-accelerated versions for faster inference on smaller GPUs.
- 8.3B parameters
- Text-to-video and image-to-video
- 480p/720p output; upscaling to 1080p
- Up to 121 frames (~5 s at 24 fps)
AccessOpen weights and inference code on Hugging Face/GitHub (Tencent-Hunyuan Community licence)