[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-1198":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":15,"stars7d":17,"stars30d":18,"stars90d":16,"forks30d":16,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":24,"hasPages":22,"topics":25,"createdAt":10,"pushedAt":10,"updatedAt":26,"readmeContent":27,"aiSummary":28,"trendingCount":16,"starSnapshotCount":16,"syncStatus":14,"lastSyncTime":29,"discoverSource":30},1198,"Hy3-preview","Tencent-Hunyuan\u002FHy3-preview","Tencent-Hunyuan","Hy3 preview (295B A21B), a leading reasoning and agent model in its size, with great cost efficiency","",null,"Python",373,18,2,5,0,16,52,15,3.84,"Other",false,"main",true,[],"2026-06-12 02:00:24","\u003Cp align=\"left\">\n    \u003Ca href=\"README_CN.md\">中文\u003C\u002Fa> ｜ English\n\u003C\u002Fp>\n\u003Cbr>\n\n\u003Cp align=\"center\">\n \u003Cimg src=\"assets\u002Flogo-en.png\" width=\"400\"\u002F> \u003Cbr>\n\u003C\u002Fp>\n\n\u003Cdiv align=\"center\" style=\"line-height: 1;\">\n\n\n[![License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Tencent%20Hy%20Community-blue)](#license)\n  \n[![HuggingFace](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-Tencent%20Hy-ffc107?color=ffc107&logoColor=white)](https:\u002F\u002Fhuggingface.co\u002Ftencent\u002FHy3-preview)\n  \n[![ModelScope](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FModelScope-Tencent%20Hy-624aff)](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FTencent-Hunyuan\u002FHy3-preview)\n  \n[![cnb.cool](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcnb.cool-Tencent%20Hy-blue?logoColor=white)](https:\u002F\u002Fcnb.cool\u002Fai-models\u002Ftencent\u002FHy3-preview)\n  \n[![GitCode](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitCode-Tencent%20Hy-red?logoColor=white)](https:\u002F\u002Fai.gitcode.com\u002Ftencent_hunyuan\u002FHy3-preview)\n\n\u003C\u002Fdiv>\n\n\u003Cp align=\"center\">\n    🖥️ \u003Ca href=\"https:\u002F\u002Faistudio.tencent.com\u002F\">\u003Cb>Official Website\u003C\u002Fb>\u003C\u002Fa>  |  \n    💬 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTencent-Hunyuan\u002FHy3-preview\">\u003Cb>GitHub\u003C\u002Fb>\u003C\u002Fa>\u003C\u002Fp>\n\n---\n\n## Table of Contents\n\n- [Model Introduction](#model-introduction)\n- [Highlights](#highlights)\n- [Benchmark Results](#benchmark-results)\n  - [STEM & Reasoning](#stem--reasoning)\n  - [Context Learning & Instruction Following](#context-learning--instruction-following)\n  - [Code & Agent](#code--agent)\n- [News](#news)\n- [Model Links](#model-links)\n- [Quickstart](#quickstart)\n- [Deployment](#deployment)\n  - [vLLM](#vllm)\n  - [SGLang](#sglang)\n- [Training](#training)\n- [Quantization](#quantization)\n- [License](#license)\n- [Contact Us](#contact-us)\n\n---\n\n## Model Introduction\n\n**Hy3 preview** is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team. Hy3 preview is the first model trained on our rebuilt infrastructure, and the strongest we've shipped so far. It improves significantly on complex reasoning, instruction following, context learning, coding, and agent tasks.\n\n\n| Property | Value |\n|:---|:---|\n| Architecture | Mixture-of-Experts (MoE) |\n| Total Parameters | 295B |\n| Activated Parameters | 21B |\n| MTP Layer Parameters | 3.8B |\n| Number of Layers (excluding MTP layer) | 80 |\n| Number of MTP Layers | 1 |\n| Attention Heads | 64 (GQA, 8 KV heads, head dim 128) |\n| Hidden Size | 4096 |\n| Intermediate Size | 13312 |\n| Context Length | 256K |\n| Vocabulary Size | 120832 |\n| Number of Experts | 192 experts, top-8 activated |\n| Supported Precisions | BF16 |\n\n## Highlights\n\n- **STEM & Reasoning** — Complex reasoning underpins everything else. Hy3 preview performs well on challenging STEM benchmarks like FrontierScience-Olympiad and IMOAnswerBench, and achieved excellent results in the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring '26) and the China High School Biology Olympiad (CHSBO 2025), demonstrating generalizable reasoning capacity.\n\n- **Context Learning & Instruction Following** — Real-world tasks require the ability to parse messy, lengthy contexts and follow complex rules. We built CL-bench and CL-bench-Life from our own business scenarios to innovatively measure context learning ability. Hy3 preview exhibits solid gains in both context learning and instruction following capabilities.\n\n- **Code & Agent** — Coding and agents saw the biggest gains. With a rebuilt RL infrastructure and larger-scale training tasks, we posted competitive scores across mainstream coding agent benchmarks (SWE-bench Verified, Terminal-Bench 2.0) and search agent benchmarks (BrowseComp, WideSearch).\n\n## Benchmark Results\n\n### Pre-trained Model Performance\n\n| Category | Benchmark (Metric) | # Shots | Kimi-K2 BASE | DeepSeek-V3 BASE | GLM-4.5 BASE | Hy3 preview-Base |\n|---|---|---|---|---|---|---|\n| | #ActivatedParams | - | 32B | 37B | 32B | 21B |\n| | #TotalParams | - | 1043B | 671B | 355B | 295B |\n| **English** | MMLU | 5-shot | **88.24** | 87.68 | 87.73 | 87.42 |\n| | MMLU-Pro | 5-shot | **65.98** | 63.98 | 63.67 | 65.76 |\n| | MMLU-Redux | 5-shot | **87.18** | 86.81 | 86.56 | 86.86 |\n| | ARC-Challenge | 0-shot | **96.66** | 94.65 | 96.32 | 95.99 |\n| | DROP | 5-shot | 86.40 | **86.50** | 82.90 | 85.50 |\n| | PIQA | 4-shot | **84.93** | 84.22 | 84.71 | 84.39 |\n| | SuperGPQA | 5-shot | 51.10 | 46.17 | 49.64 | **51.60** |\n| | SimpleQA | 5-shot | **34.37** | 26.15 | 29.26 | 26.47 |\n| **Code** | MBPP-plus | 3-shot | **81.35** | 75.47 | 78.05 | 78.71 |\n| | CRUXEval-I | 3-shot | 68.01 | 67.79 | 68.51 | **71.19** |\n| | CRUXEval-O | 3-shot | 69.62 | **71.00** | 67.75 | 68.38 |\n| | LiveCodeBench-v6 | 1-shot | 30.86 | 29.31 | 27.43 | **34.86** |\n| **Math** | GSM8K | 4-shot | 93.46 | 88.15 | 90.06 | **95.37** |\n| | MATH | 4-shot | 71.20 | 59.37 | 61.00 | **76.28** |\n| | CMath | 4-shot | 90.83 | 85.50 | 89.33 | **91.17** |\n| **Chinese** | C-Eval | 5-shot | **91.51** | 90.35 | 85.84 | 89.80 |\n| | CMMLU | 5-shot | **90.72** | 87.90 | 86.46 | 89.61 |\n| | Chinese-simpleQA | 5-shot | **74.58** | 68.72 | 68.49 | 69.73 |\n| **Multilingual** | MMMLU | 5-shot | 77.63 | 79.54 | 79.26 | **80.15** |\n| | INCLUDE | 5-shot | 75.66 | 77.86 | 76.27 | **78.64** |\n\n### Instruct Model Performance\n\n#### STEM & Reasoning\n\nComplex reasoning underpins everything else. Hy3 preview performs well on challenging STEM benchmarks like FrontierScience-Olympiad and IMOAnswerBench. It also achieved excellent results in the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring '26) and the China High School Biology Olympiad (CHSBO 2025), demonstrating a high degree of generalizable reasoning capacity.\n\n\u003Cp align=\"center\">\u003Cimg src=\"assets\u002Fbench_stem.jpg\" width=\"800\" alt=\"STEM & Reasoning benchmarks\"\u002F>\u003C\u002Fp>\n\n#### Context Learning & Instruction Following\n\nReal-world tasks require the ability to parse messy, lengthy contexts and follow complex rules. We built CL-bench and CL-bench-Life from our own business scenarios to innovatively measure context learning ability. Hy3 preview exhibits solid gains in both context learning and instruction following capabilities.\n\n\u003Cp align=\"center\">\u003Cimg src=\"assets\u002Fbench_context.jpg\" width=\"800\" alt=\"Context Learning & Instruction Following benchmarks\"\u002F>\u003C\u002Fp>\n\n#### Code & Agent\n\nCoding and agents saw the biggest gains. With a rebuilt RL infrastructure and larger-scale training tasks, we posted competitive scores across mainstream coding agent benchmarks (SWE-bench Verified, Terminal-Bench 2.0) and search agent benchmarks (BrowseComp, WideSearch).\n\n\u003Cp align=\"center\">\u003Cimg src=\"assets\u002Fbench_agent_overview_v3.jpg\" width=\"800\" alt=\"Agent benchmarks overview\"\u002F>\u003C\u002Fp>\n\nCoding is about whether a model can execute in a development environment. Search is about whether it can find and combine information from the open web. Both matter for complex agent scenarios like OpenClaw. Hy3 preview scores well on ClawEval and WildClawBench — a sign that its agent capabilities are becoming practical.\n\n\u003Cp align=\"center\">\u003Cimg src=\"assets\u002Fbench_claw_agent.png\" width=\"800\" alt=\"Claw Agent benchmarks\"\u002F>\u003C\u002Fp>\n\nBeyond public benchmarks, we built internal evaluation sets to test the model in real development scenarios. On Hy-Backend (backend-focused tasks), Hy-Vibe Bench (real-user dev workflows), and Hy-SWE Max, Hy3 preview scores competitively against other open-source models.\n\n\u003Cp align=\"center\">\u003Cimg src=\"assets\u002Fbench_claw_agent2.jpg\" width=\"800\" alt=\"Internal benchmarks\"\u002F>\u003C\u002Fp>\n\n## News\n\n\n* **[2026-04-23]** 🔥 We open-source **Hy3 preview** model weights on [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Ftencent\u002FHy3-preview), [ModelScope](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FTencent-Hunyuan\u002FHy3-preview), and [GitCode](https:\u002F\u002Fai.gitcode.com\u002Ftencent_hunyuan\u002FHy3-preview).\n\n## Model Links\n\n\n| Model Name | Description | Hugging Face | ModelScope | GitCode |\n|:---|:---|:---:|:---:|:---:|\n| Hy3 preview | Instruct model | 🤗 [Model](https:\u002F\u002Fhuggingface.co\u002Ftencent\u002FHy3-preview) | [Model](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FTencent-Hunyuan\u002FHy3-preview) | [Model](https:\u002F\u002Fai.gitcode.com\u002Ftencent_hunyuan\u002FHy3-preview) |\n| Hy3 preview-Base | Pre-trained base model | 🤗 [Model](https:\u002F\u002Fhuggingface.co\u002Ftencent\u002FHy3-preview-Base) | [Model](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FTencent-Hunyuan\u002FHy3-preview-Base) | [Model](https:\u002F\u002Fai.gitcode.com\u002Ftencent_hunyuan\u002FHy3-preview-Base) |\n\n## Quickstart\n\nDeploy Hy3 preview with [vLLM](#vllm) or [SGLang](#sglang) first, then call the OpenAI-compatible API:\n\n```python\nfrom openai import OpenAI\n\nclient = OpenAI(base_url=\"http:\u002F\u002Flocalhost:8000\u002Fv1\", api_key=\"EMPTY\")\n\nresponse = client.chat.completions.create(\n    model=\"hy3-preview\",\n    messages=[\n        {\"role\": \"user\", \"content\": \"Hello! Can you briefly introduce yourself?\"},\n    ],\n    temperature=0.9,\n    top_p=1.0,\n    # reasoning_effort: \"no_think\" (default, direct response), \"low\", \"high\" (deep chain-of-thought)\n    extra_body={\"chat_template_kwargs\": {\"reasoning_effort\": \"no_think\"}},\n)\nprint(response.choices[0].message.content)\n```\n\n> **Recommended parameters**: `temperature=0.9`, `top_p=1.0`.\n>\n> **Reasoning mode**: Set `reasoning_effort` to `\"high\"` for complex tasks (math, coding, reasoning) or `\"no_think\"` for direct responses.\n\nSee the [Deployment](#deployment) section below for how to start the API server.\n\n## Deployment\n\nHy3-preview has 295B parameters in total. To serve it on 8 GPUs, we recommend using H20-3e or other GPUs with larger memory capacity.\n\n### vLLM\n\nBuild vLLM from source:\n```bash\nuv venv --python 3.12 --seed --managed-python\nsource .venv\u002Fbin\u002Factivate\ngit clone https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm.git\ncd vllm\nuv pip install --editable . --torch-backend=auto\n```\n\nStart the vLLM server with MTP enabled:\n\n```bash\nvllm serve tencent\u002FHy3-preview \\\n  --tensor-parallel-size 8 \\\n  --speculative-config.method mtp \\\n  --speculative-config.num_speculative_tokens 1 \\\n  --tool-call-parser hy_v3 \\\n  --reasoning-parser hy_v3 \\\n  --enable-auto-tool-choice \\\n  --served-model-name hy3-preview\n```\n\n### SGLang\n\nBuild SGLang from source:\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\ncd sglang\npip3 install pip --upgrade\npip3 install \"transformers>=5.6.0\"\npip3 install -e \"python\"\n```\n\nLaunch SGLang server with MTP enabled:\n\n```bash\npython3 -m sglang.launch_server \\\n  --model tencent\u002FHy3-preview \\\n  --tp 8 \\\n  --tool-call-parser hunyuan \\\n  --reasoning-parser hunyuan \\\n  --speculative-num-steps 1 \\\n  --speculative-eagle-topk 1 \\\n  --speculative-num-draft-tokens 2 \\\n  --speculative-algorithm EAGLE \\\n  --served-model-name hy3-preview\n```\n\n## Training\n\nHy3 preview provides a complete model training pipeline, supporting both full fine-tuning and LoRA fine-tuning, with DeepSpeed ZeRO configurations and LLaMA-Factory integration.\n\nFor detailed training documentation, please refer to: [Training Guide](.\u002Ftrain\u002FREADME.md)\n\n## Quantization\n\nWe provide [AngelSlim](https:\u002F\u002Fgithub.com\u002Ftencent\u002FAngelSlim), a more accessible, comprehensive, and efficient toolkit for large model compression. AngelSlim supports a comprehensive suite of compression tools for large-scale multimodal models, including common quantization algorithms, low-bit quantization, and speculative sampling.\n\n## License\n\n\nHy3 preview is released under the **Tencent Hy Community License Agreement**. See [LICENSE](.\u002FLICENSE) for details.\n\n## Contact Us\n\nIf you would like to leave a message for our R&D and product teams, welcome to contact us. You can also reach us via email:\n\n📧 **hunyuan_opensource@tencent.com**\n\n---\n\n\u003Cp align=\"center\">\n  \u003Ci>Hy3 preview is developed by the Tencent Hy Team.\u003C\u002Fi>\n\u003C\u002Fp>\n","Hy3 preview 是由腾讯混元团队开发的一款295B参数的混合专家模型，激活参数为21B。该模型在复杂推理、指令执行、上下文学习、编程和代理任务等方面表现出色，特别是在STEM领域的挑战性基准测试中取得了优异成绩。采用Mixture-of-Experts架构，支持BF16精度计算，具有80层主网络结构及单层MTP层，适用于需要高效处理大量数据并进行复杂推理的应用场景，如科研、教育和技术开发等领域。","2026-06-11 02:42:15","CREATED_QUERY"]