Tencent Hy3 模型概覽

Tencent Hy3 Model Overview

Executive Summary

Tencent 已發佈 Hy3 的完整版本,這是一款混合專家模型 (MoE),旨在競爭中階前沿模型領域。Hy3 特別針對代理任務 (agentic tasks)、工具使用和本地部署進行了優化,使其成為比 GLM 5.2 等大型模型更具硬體效率的替代方案,同時在推理和減少幻覺方面保持高性能。

Model Architecture and Specifications

Hy3 被構建為一個大規模混合專家模型,重點在於平衡效能與效率。

  • Parameter Count: 295 billion total parameters.
  • Active Parameters: 21 billion active parameters per token.
  • Speculative Decoding: Includes a 3.8 billion parameter speculative decoding model to increase inference speed.
  • Context Window: 256K tokens.

Performance and Benchmarks

Hy3 被定位為「中階」模型,目標是填補小型本地模型與大型專有前沿模型之間的空間。

Agentic Tasks and Tool Use

Hy3 在代理工作流 (agentic workflows) 中表現出色,特別是在工具調用 (tool calling) 和輸出格式化方面。在測試中,該模型展現了高度的熟練度:

  • Repeated Tool Calls: Successfully handling multiple sequential tool calls.
  • Pagination: Managing long-running pagination across twelve different tools.
  • Error Recovery: Demonstrating resilience by attempting retries when tool calls return errors rather than giving up.
  • Noise Filtering: Identifying relevant information from API responses without being distracted by irrelevant data.

Comparison with GLM 5.2

雖然 Hy3 功能強大,但其目的並非取代所有高端模型。具體而言,GLM 5.2 在代理編碼任務 (agentic coding tasks) 中通常優於 Hy3。然而,Hy3 的規模顯著較小 (遠大於 GLM 5.2 的一半),使其在本地託管和在私有硬體上進行微調變得更加可行,而不需要 B200 GPU 集群。

Reliability and Hallucinations

Tencent 已投入大量精力於後訓練 (post-training) 和數據清洗,以提高可靠性。與其預覽版本相比,完整的 Hy3 模型將常識錯誤率和幻覺率都降低了一半。

Capabilities and Testing Results

Reasoning and Chain-of-Thought

Hy3 利用了長鏈式思考 (CoT) 過程。在邏輯謎題測試中,該模型生成了大量的「思考」標記 (tokens),以在得出解決方案之前驗證其步驟。這種內部推理的品質被認為很高,甚至可能超過某些其他開源模型。

Creative and Technical Generation

  • SVG Generation: The model can generate complex SVG code, such as a detailed pelican on a bicycle, showing significant improvement over the preview version.
  • HTML/CSS: Hy3 is capable of producing polished, functional website layouts, including opt-in forms and integrated images.
  • Long-form Content: In a 5,000-word essay test, the model produced a structured outline (acting as an agentic planning step) and generated approximately 2,500 to 3,000 words, noting its own constraints regarding single-block generation.

Deployment and Accessibility

Hy3 目前可透過 OpenRouter 進行測試。由於其規模和架構,它被視為那些希望擁有完全封閉、本地模型,並能在可控硬體上針對特定企業用途進行微調的公司之強大候選方案。

Sources