AI News AI资讯 4d ago Updated 3d ago 更新于 3天前 51

Tencent releases Hy3 open-source model that allegedly matches models up to five times its active size 腾讯发布开源模型Hy3,据称性能匹敌参数量大五倍的模型

Tencent released Hy3, an open-source Mixture-of-Experts (MoE) model with 295 billion total parameters but only 21 billion active parameters per inference step. The model features a 3.8 billion parameter Multi-Token Prediction (MTP) layer and supports a context window of up to 256,000 tokens. In blind evaluations by 270 experts, Hy3 achieved a score of 2.67/4, outperforming GLM-5.1 (2.51/4), while reducing hallucination rates from 12.5% to 5.4%. Hy3 is available under the Apache 2.0 license on ma 腾讯发布开源大模型 Hy3,采用 MoE 架构,总参数量 2950 亿,激活参数仅 210 亿,并额外包含 38 亿参数的 MTP 层。 官方宣称 Hy3 性能可匹敌参数量为其两到五倍的模型,并在 270 位专家盲测中得分 2.67/4,超越 GLM-5.1。 模型支持高达 256,000 token 的上下文长度,内部测试显示幻觉率从 12.5% 降至 5.4%。 Hy3 以 Apache 2.0 协议在 Hugging Face、ModelScope 和 GitHub 开源,并提供 FP8 量化版本。 该模型已集成至微信、元宝、WorkBuddy 及《流放之路2》游戏助手等腾讯内部产品中。

75
Hot 热度
70
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • Tencent released Hy3, an open-source Mixture-of-Experts (MoE) model with 295 billion total parameters but only 21 billion active parameters per inference step.
  • The model features a 3.8 billion parameter Multi-Token Prediction (MTP) layer and supports a context window of up to 256,000 tokens.
  • In blind evaluations by 270 experts, Hy3 achieved a score of 2.67/4, outperforming GLM-5.1 (2.51/4), while reducing hallucination rates from 12.5% to 5.4%.
  • Hy3 is available under the Apache 2.0 license on major platforms, including an FP8-quantized version, and is already integrated into Tencent’s ecosystem products like WeChat and WorkBuddy.

Why It Matters

This release demonstrates the practical viability of highly sparse MoE architectures for achieving frontier-level performance with significantly reduced computational costs during inference. By making such a large-scale model openly available under a permissive license, Tencent lowers the barrier for developers to experiment with high-capability models that were previously accessible only through proprietary APIs.

Technical Details

  • Architecture: Utilizes a Mixture-of-Experts (MoE) design with 295 billion total parameters, activating only 21 billion parameters per token generation to optimize efficiency.
  • Enhancements: Incorporates a 3.8 billion parameter Multi-Token Prediction (MTP) layer to improve generation speed and accuracy.
  • Context Capacity: Supports long-context understanding with a maximum sequence length of 256,000 tokens.
  • Performance Metrics: Internal testing confirmed a hallucination reduction from 12.5% to 5.4%, and expert blind tests rated it superior to competitors like GLM-5.1.
  • Availability: Distributed via Hugging Face, ModelScope, and GitHub under Apache 2.0, with an FP8-quantized variant provided for broader hardware compatibility.

Industry Insight

The aggressive open-sourcing of a model with such a high parameter count suggests a strategic move to establish Hy3 as a foundational standard in the open-source AI community, potentially challenging the dominance of closed-source proprietary models. Practitioners should evaluate the FP8 quantization for deployment on consumer-grade GPUs, as this could enable high-performance inference on hardware previously considered insufficient for models of this scale.

TL;DR

  • 腾讯发布开源大模型 Hy3,采用 MoE 架构,总参数量 2950 亿,激活参数仅 210 亿,并额外包含 38 亿参数的 MTP 层。
  • 官方宣称 Hy3 性能可匹敌参数量为其两到五倍的模型,并在 270 位专家盲测中得分 2.67/4,超越 GLM-5.1。
  • 模型支持高达 256,000 token 的上下文长度,内部测试显示幻觉率从 12.5% 降至 5.4%。
  • Hy3 以 Apache 2.0 协议在 Hugging Face、ModelScope 和 GitHub 开源,并提供 FP8 量化版本。
  • 该模型已集成至微信、元宝、WorkBuddy 及《流放之路2》游戏助手等腾讯内部产品中。

为什么值得看

Hy3 展示了通过高效 MoE 架构和 MTP 技术实现高参数效率与低幻觉率的最新进展,为行业提供了极具竞争力的开源基线模型。其开源策略及在腾讯核心产品中的落地应用,揭示了大型科技公司如何通过自研模型优化用户体验并降低推理成本。

技术解析

  • 架构设计:采用混合专家(MoE)架构,总参数量达 2950 亿,但单次推理仅激活 210 亿参数,显著降低了计算开销。此外,引入了 38 亿参数的多令牌预测(MTP)层以加速推理或提升效果。
  • 性能表现:在 270 名专家的盲测评估中,Hy3 获得 2.67 分(满分 4 分),优于智谱 GLM-5.1 的 2.51 分。官方声称其性能等效于参数量大 2-5 倍的密集模型。
  • 可靠性与上下文:支持 256K 长上下文窗口。内部测试数据显示,经过优化后模型的幻觉率从 12.5% 大幅降低至 5.4%,提升了输出的准确性。
  • 开源与部署:模型以 Apache 2.0 许可证开源,提供标准版及 FP8 量化版本,便于不同硬件环境的部署。计划支持 OpenRouter 和 Cline 等平台,并已嵌入微信、元宝等多个腾讯生态应用。

行业启示

  • MoE 架构成为主流优化路径:Hy3 的高参数总量与低激活参数比再次验证了 MoE 架构在平衡模型能力与推理成本方面的优势,未来更多模型将趋向于此设计。
  • 幻觉抑制是关键竞争点:将幻觉率从 12.5% 降至 5.4% 是显著的技术突破,表明降低幻觉已成为衡量企业级模型实用性的核心指标之一。
  • 开源与内部生态闭环并行:腾讯在保持核心产品(如微信)使用自研模型的同时开放基础版本,既推动了行业标准建立,又确保了自身生态的技术壁垒和数据飞轮效应。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Research 科学研究 Product Launch 产品发布