Tencent’s open-weight AI model family. April 2026 was the family’s strongest month on record: HY-World 2.0 open-sourced 3D world generation (April 16), Hy3-preview (April 23) shipped a 295B MoE LLM that rivals Qwen3.6 and DeepSeek V3 on reasoning and coding, HY-OmniWeaving (April 3) consolidated eight video tasks behind a single HunyuanVideo-1.5 backbone, and HY-Embodied-0.5 extended the family into image-text-to-text embodied AI. By the June 2 refresh, HY-World 2.0’s image-to-3DWorld generation code was also released. Most weights ship on Hugging Face under permissive licenses; hosted API access runs through Tencent Cloud.
System Verdict
Pick Hunyuan if you need open-source 3D world generation, embodied AI, or a frontier-grade open MoE LLM. HY-World 2.0 remains category-leading among open-source options, with editable output that imports cleanly into Unity, Unreal Engine, and Isaac. Hy3-preview (April 23, 2026) is now genuinely competitive with Qwen3.6 Plus and DeepSeek V3 on agentic coding and STEM reasoning, with a 256K context and only 21B activated parameters per token. HY-OmniWeaving unifies eight video-generation tasks behind one model, narrowing the gap with closed-source video frontiers for self-hosted use.
Skip it if you want English-first consumer chat or production-grade commercial video output. Hy3-preview is a research-friendly weights drop; product polish still trails Qwen Chat and GLM for bilingual chat workflows. HY-OmniWeaving improves on HunyuanVideo but still trails Kling, Veo 3.1, and Runway Gen-4 on commercial output quality.
Where Hunyuan wins uniquely: Open-source 3D world generation plus open-weight frontier reasoning plus unified open video, all under one organization. For game developers, AR/VR teams, robotics researchers, simulation engineers, and open-source-first R&D labs, Hunyuan is now the default starting point.
Key Facts
| Parent company | Tencent (00700.HK) |
| Flagship LLM (June 2026) | Hy3-preview · 295B MoE / 21B activated · 256K context · released April 23, 2026 |
| Hy3-preview benchmarks | SWE-bench Verified 74.4% · Terminal-Bench 2.0 54.4% · GPQA Diamond 87.2% |
| Hy3-preview architecture | 192 experts, top-8 activated, BF16, open-weights on Hugging Face / ModelScope / GitCode |
| Flagship 3D world model | HY-World 2.0 · released April 16, 2026 · open-source |
| HY-World 2.0 output | 3D Gaussian Splattings, editable meshes, point clouds; image-to-3DWorld generation code released |
| HY-World 2.0 imports to | Unity, Unreal Engine, Nvidia Isaac |
| HY-World 2.0 architecture | 4-stage pipeline: HY-Pano 2.0, WorldNav, WorldStereo 2.0, WorldMirror 2.0 |
| WorldMirror 2.0 size | ~1.2B parameters |
| HY-World 2.0 resolution | 50K to 500K pixels, flexible-resolution inference |
| Unified video model | HY-OmniWeaving · April 3, 2026 · 8 unified video tasks, built on HunyuanVideo-1.5 |
| OmniWeaving features | Thinking Mode, Hidden States DeepStacking, MLLM + MMDiT + VAE architecture |
| Embodied AI | HY-Embodied-0.5 · 4B params, image-text-to-text |
| Previous LLM family | Hunyuan-Large (389B MoE), Hunyuan-Turbo, Hunyuan-Standard |
| 3D asset model | Hunyuan3D (text-to-3D asset generation) |
| Hosted API | Tencent Cloud Hunyuan API |
| Open-weight hub | Hugging Face tencent org (119 models, 400+ team members) |
When to pick Hunyuan
- Game developers generating 3D scenes. HY-World 2.0 outputs are engine-importable. Text or image input becomes a playable level in Unity or Unreal.
- Robotics and simulation teams. Isaac import path plus open-source license makes this the default for training embodied agents in procedurally generated environments. HY-Embodied-0.5 adds a dedicated image-text-to-text model for the same audience.
- AR and VR content pipelines. Gaussian Splatting output is first-class in modern XR workflows; meshes work for mobile AR and WebXR.
- Open-weight agentic coding. Hy3-preview’s 74.4% SWE-bench Verified score puts it within reach of closed-source frontier LLMs while running fully open under permissive license terms.
- Chinese-market LLM deployment. Hy3-preview and Hunyuan-Large are tuned for Chinese language quality and integrate natively with Tencent Cloud services popular in China.
- Unified open-weight video pipelines. HY-OmniWeaving handles eight generation tasks in one model, useful for teams that want to self-host instead of paying per-clip on Kling or Veo 3.1.
- Researchers working on world models or large MoE LLMs. Open code plus weights plus technical reports make HY-World 2.0 and Hy3-preview credible research baselines.
When to pick something else
- Production video generation: Kling, Veo 3, Runway, Hailuo for commercial video output.
- English-first chat: ChatGPT, Claude, Gemini, or Qwen for multilingual.
- 3D character generation specifically: Rodin, Meshy, Tripo3D focus on single-asset 3D generation.
- Closed-source world modeling at commercial grade: Marble (closed, similar capability to HY-World 2.0 but licensed).
Pricing
Hunyuan models are distributed primarily as open weights through Tencent Cloud is pay-as-you-go and varies by model.
| Surface | Price |
|---|---|
| Hy3-preview (open weights) | Free · Hugging Face |
| HY-World 2.0 (open weights) | Free · Hugging Face |
| HY-OmniWeaving (open weights) | Free · Hugging Face |
| HY-Embodied-0.5 (open weights) | Free · Hugging Face |
| Hunyuan-Large (open weights) | Free · Hugging Face |
| HunyuanVideo (open weights) | Free · Hugging Face |
| Hunyuan3D (open weights) | Free · Hugging Face |
| Tencent Cloud hosted API | Pay-as-you-go (see Tencent Cloud pricing) |
Self-hosted runs under the respective license terms. Commercial use allowed on most releases; verify the specific license in each repo.
Recent developments
- April 23, 2026: Hy3-preview released. 295B MoE / 21B activated with a 256K context and 192-expert top-8 routing. Benchmarks land at 74.4% on SWE-bench Verified, 54.4% on Terminal-Bench 2.0, and 87.2% on GPQA Diamond, putting it in the same conversation as Qwen3.6 Plus and DeepSeek V3 for open-weight agentic coding and STEM reasoning. Available on Hugging Face, ModelScope, and GitCode.
- April 16, 2026: Tencent open-sources Hunyuan World 2.0 with editable 3DGS plus mesh output and Unity, Unreal, and Isaac import. Cross-reference Google’s Gemini 3D Models preview from April 10, which is closed-source and narrower in scope.
- May 18, 2026: HY-World 2.0 image-to-3DWorld generation code was released, reducing the launch-time gap between paper/model-card claims and usable open-source implementation.
- April 3, 2026: HY-OmniWeaving code and weights ship. Unified video model on the HunyuanVideo-1.5 backbone covering T2V, I2V, key-frames-to-video, reference-to-video, video editing, text-image-video-to-video, compositional multi-image-to-video, and reasoning-augmented generation.
- Late March / April 2026: HY-Embodied-0.5 extends the family into embodied AI with a 4B image-text-to-text model targeting robotics and on-device perception.
Failure modes
- Documentation lag in English. Primary docs are in Chinese first. English translations trail the original by days to weeks.
- Hy3-preview is a research preview. Branded “preview” rather than a stable release; expect tokenizer, prompt-template, and serving-stack updates before a 1.0 cut.
- Open-source release still needs repo-by-repo verification. HY-World 2.0 image-to-3DWorld code is now released, but teams should still check task-specific scripts, licenses, checkpoints, and dependency health before building a production pipeline.
- Video generation still trails commercial frontier. HY-OmniWeaving is a meaningful jump over HunyuanVideo for open-weight users but does not match Kling, Veo 3.1, or Runway Gen-4 on commercial output quality.
- Hosted API is Chinese-cloud. Enterprise buyers in regulated Western markets need to evaluate Tencent Cloud data residency posture. Self-hosting open weights is the workaround.
- LLM product polish trails Qwen and DeepSeek. Hy3-preview narrows the gap on benchmarks but Qwen Chat and DeepSeek consumer apps remain easier first stops for non-developer users.
- No equivalent to ChatGPT plugin marketplace. Hunyuan is model-first, not a product ecosystem.
Against the alternatives (3D world modeling)
| HY-World 2.0 | Marble (closed) | HunyuanWorld 1.0 | |
|---|---|---|---|
| License | Open source | Closed | Open source |
| Output | 3DGS + meshes (editable) | Similar | Video only |
| Engine import | Unity, Unreal, Isaac | Proprietary | No direct import |
| Resolution | 50K-500K pixels | Similar | N/A |
| Best for | Open-source game dev + robotics | Enterprise 3D workflows | Walkthrough video |
Methodology
Produced by the aipedia.wiki editorial pipeline. Last verified 2026-06-12 against Tencent-Hunyuan/HY-World-2.0 GitHub, the Hugging Face tencent org, the Hy3-preview model card, the HY-OmniWeaving model card, and Tencent Cloud Hunyuan.
FAQ
What is HY-World 2.0? Tencent’s open-source 3D world model, released April 16, 2026. Generates editable 3D worlds (Gaussian Splattings + meshes) from text, images, or video inputs. Outputs import directly into Unity, Unreal Engine, and Nvidia Isaac. Architecture is a 4-stage pipeline: HY-Pano 2.0 → WorldNav → WorldStereo 2.0 → WorldMirror 2.0.
How is HY-World 2.0 different from 1.0? HY-World 1.0 (called WorldPlay + WorldCompass) generated pixel-level walkthrough videos of scenes. You got video frames, not reusable assets. HY-World 2.0 generates persistent, editable 3D assets you can open in a game engine and modify.
Is Hunyuan really free? Most model weights are open-source on Hugging Face under permissive licenses (check each repo for specific terms). Self-hosting is free. Tencent Cloud hosted API is pay-as-you-go.
Does Hunyuan have a consumer chat product? Tencent runs a Hunyuan chat surface at hunyuan.tencent.com, primarily for Chinese-market users. English coverage is limited compared to Qwen or DeepSeek.
How does Hunyuan-Large compare to Qwen3 or DeepSeek V3? Hunyuan-Large is a 389B MoE and is capable but trails Qwen3.6 Plus and DeepSeek V3 on most independent benchmarks. Hy3-preview (April 23, 2026) materially closes that gap with 74.4% on SWE-bench Verified and 87.2% on GPQA Diamond, putting it back in the open-weight frontier conversation. Hunyuan’s broader advantage is still the multi-modality family spanning LLMs, 3D world models, video, and embodied AI.
What is Hy3-preview? Tencent’s April 23, 2026 open-weight MoE LLM. 295B total parameters with 21B activated per token, 192 experts top-8 routing, 256K context, BF16 precision. Strong on STEM, agentic coding (74.4% SWE-bench Verified, 54.4% Terminal-Bench 2.0), and instruction following. Released on Hugging Face, ModelScope, and GitCode. Branded “preview” rather than a stable 1.0.
What is HY-OmniWeaving? A unified video generation model released April 3, 2026 on the HunyuanVideo-1.5 backbone. Handles eight tasks in a single model: text-to-video, image-to-video, key-frames-to-video, reference-to-video, video editing, text-image-video-to-video, compositional multi-image-to-video, and reasoning-augmented generation. Adds a “Thinking Mode” where an MLLM reasons over user intent before generation.
Sources
Related
- Category: AI Chatbots · AI Video · AI Image
- See also: Qwen · DeepSeek · GLM · Kling · Rodin
- News: Tencent Open-Sources Hunyuan 3D World Model 2.0 · Google Gemini 3D Models preview (April 10, 2026)
