AI News AI资讯 3h ago Updated 1h ago 更新于 1小时前 48

Introducing Hy4 Preview 腾讯发布Hy4 Preview

Tencent released Hy4 Preview, an open-weight text-only LLM with 770B total parameters and 49B active parameters, significantly larger than its predecessor Hy3 The model supports a 1M token context window and is available as a 1.56TB download on Hugging Face Hy4 features two reasoning modes: "high" (default, chain-of-thought enabled) and "no_think" (reasoning disabled) This represents a major scaling leap from Hy3 (295B total / 21B active, 256K context, 598GB) released just one month prior The mo 腾讯发布Hy4 Preview开源大模型,总参数770B,激活参数49B,为当前最大规模开源模型之一 上下文窗口达1M token,较前代Hy3提升近4倍(256K→1M) 模型仅支持文本输入(无视觉能力),提供high/no_think两种推理模式 模型权重1.56TB托管于Hugging Face,相比Hy3的598GB增长约2.6倍 推理过程使用略微截断的英语,推测为提升token效率的优化策略

72
Hot 热度
68
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Tencent released Hy4 Preview, an open-weight text-only LLM with 770B total parameters and 49B active parameters, significantly larger than its predecessor Hy3
  • The model supports a 1M token context window and is available as a 1.56TB download on Hugging Face
  • Hy4 features two reasoning modes: "high" (default, chain-of-thought enabled) and "no_think" (reasoning disabled)
  • This represents a major scaling leap from Hy3 (295B total / 21B active, 256K context, 598GB) released just one month prior
  • The model is text-input only with no vision capabilities

Why It Matters

Tencent's rapid iteration from Hy3 to Hy4 in under a month signals intensifying competition in the open-weight large language model space, particularly from Chinese AI labs. The 1M token context window positions Hy4 as a strong contender for long-context applications such as document analysis, codebase understanding, and extended reasoning tasks. The dual reasoning-mode design offers practitioners flexibility between performance and latency/cost trade-offs.

Technical Details

  • Architecture: Mixture-of-Experts (MoE) design with 770B total parameters and 49B active parameters per token, enabling efficient inference despite massive scale
  • Context Window: 1,000,000 tokens, a 4x increase over Hy3's 256K context, supporting extremely long document and conversation processing
  • Reasoning Modes: Two explicit modes via reasoning_effort parameter — "high" (default, full chain-of-thought reasoning) and "no_think" (direct response, no visible reasoning trace)
  • Model Size: 1.56TB checkpoint on Hugging Face, up from Hy3's 598GB, reflecting the substantial parameter and context scaling
  • Input Modality: Text-only input with no vision capabilities, distinguishing it from multimodal competitors
  • Chat Template: Uses Jinja2-based chat templates with strict validation on reasoning effort values, raising exceptions for invalid or missing inputs

Industry Insight

  • The aggressive monthly iteration cycle (Hy3 in July → Hy4 in August) suggests Chinese AI labs are prioritizing rapid open-weight releases to capture developer mindshare, potentially pressuring Western labs to accelerate their own open model roadmaps
  • The 1M token context window at open weight is a significant milestone for enterprise applications involving long documents, legal contracts, and full-codebase analysis — practitioners should evaluate Hy4 for long-context workloads where proprietary models previously dominated
  • The explicit "no_think" mode reflects an emerging industry pattern: offering users direct control over reasoning overhead, balancing quality against speed and cost, which is increasingly important for production deployment scenarios

TL;DR

  • 腾讯发布Hy4 Preview开源大模型,总参数770B,激活参数49B,为当前最大规模开源模型之一
  • 上下文窗口达1M token,较前代Hy3提升近4倍(256K→1M)
  • 模型仅支持文本输入(无视觉能力),提供high/no_think两种推理模式
  • 模型权重1.56TB托管于Hugging Face,相比Hy3的598GB增长约2.6倍
  • 推理过程使用略微截断的英语,推测为提升token效率的优化策略

为什么值得看

腾讯Hy4的发布标志着开源大模型参数规模再次突破,770B总参数/49B激活参数的MoE架构为行业树立了新的规模标杆。1M超长上下文窗口使模型在处理长文档、代码库等复杂任务时具备更强能力,对需要长程依赖的应用场景具有重要参考价值。

技术解析

  • 模型架构:采用MoE(混合专家)架构,770B总参数中仅激活49B,兼顾推理效率与模型容量,参数量较Hy3的295B/21B实现大幅提升
  • 上下文能力:支持1M token超长上下文窗口,较Hy3的256K提升近4倍,可处理超长文本、完整代码库或长篇文档
  • 推理模式:提供"high"(默认)和"no_think"两种推理努力级别,允许用户根据任务复杂度灵活选择是否启用思维链
  • 模型规格:纯文本输入模型(无视觉模态),权重文件1.56TB,托管于Hugging Face开源
  • 推理优化:从chat_template和推理trace可见,模型在hidden reasoning阶段使用略微截断的英语,推测为减少token消耗、提升推理效率的设计

行业启示

  • 开源大模型军备竞赛持续升级,腾讯通过Hy4在参数规模和上下文长度上追赶头部闭源模型,开源生态竞争力进一步增强
  • 超长上下文(1M token)正成为大模型标配能力,企业应评估自身业务对长程依赖的需求,优先选择支持超长上下文的模型
  • 推理模式的可配置化(high/no_think)反映了行业对推理成本与效果的平衡探索,未来模型服务应提供灵活的推理策略选项

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 Product Launch 产品发布 Training 训练 Inference 推理