国产98在线 | 传媒麻豆,久久―日本道色综合久久,久久久精品久久久久久久久久久,国产成人精品免费久久久久,五月综合色婷婷影院在线观看,久久久亚洲精品视频

  • 虎科技 - 領先的互聯網科技媒體

阿里發布4B小模型Qwen3系列:性能超GPT-4.1-nano,端側部署更友好

   時間:2025-08-08 07:59 來源:ITBEAR作者:沈瑾瑜

近日,通義千問在其技術更新中揭曉了兩款全新的小型語言模型——Qwen3-4B-Instruct-2507與Qwen3-4B-Thinking-2507,這兩款模型以其精簡的尺寸和卓越的性能引起了業界的廣泛關注。

據悉,Qwen3-4B系列的兩款新模型在設計上充分考慮了端側硬件的部署需求,旨在為用戶提供更加高效、靈活的AI解決方案。其中,Qwen3-4B-Thinking-2507在推理能力上展現出了與中尺寸模型相當的實力,而Qwen3-4B-Instruct-2507則在知識、推理、編程、對齊及agent能力等多個方面全面超越了閉源的GPT-4.1-nano模型。

這兩款新模型的上下文理解能力得到了顯著提升,擴展至256K,使其能夠處理更長的文本,并支持復雜的文檔分析、長篇內容生成及跨段落推理等應用場景。在性能測試中,Qwen3-4B-Thinking-2507在復雜問題推理、數學能力、代碼能力及多輪函數調用等方面均表現出色,而Qwen3-4B-Instruct-2507則在Arena-Hard v2基準測試中取得了43.4分的高分。

值得注意的是,Qwen3-4B-Instruct-2507在非推理領域同樣展現出了強大的實力,全面超越了GPT-4.1-nano模型,并且其性能已接近中等規模的Qwen3-30B-A3B(non-thinking)模型。該模型覆蓋了更廣泛的語言知識,與人類偏好在主觀和開放性任務中的對齊性得到增強,能夠提供更貼近用戶需求的答復。

在推理領域,Qwen3-4B-Thinking-2507更是表現出色,其推理能力足以媲美中等模型Qwen3-30B-Thinking。特別是在數學能力的測試中,該模型以4B的參數量取得了81.3分的高分。其通用能力也得到了顯著提升,Agent分數甚至超越了更大尺寸的Qwen3-30B-Thinking模型。

通義千問此次推出的Qwen3-4B系列新模型,不僅實現了性能上的優化,更在通用能力上展現出了超越同級別模型甚至逼近中大規模模型的實力。同時,這兩款小型語言模型對端側硬件的友好度也更高,為用戶提供了更加靈活、高效的AI解決方案。這一創新舉措無疑將加速AI技術在端側設備的滲透,推動更多輕量化、場景化的智能應用落地。

 
 
更多>同類內容
推薦圖文
推薦內容
點擊排行
 
智快科技微信賬號
ITBear微信賬號

微信掃一掃
加微信拉群
電動汽車群
科技數碼群

主站蜘蛛池模板: | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |