InquirerRemulla: No truth to ‘chismis’ Romualdez can freely leave jailDaily MaverickWHAT WE’RE WATCHING: Resident Evil — a horror video game adaptation with a wicked sense of humourPunchNJC suspends three judges, disciplines six others한겨레미 데이터센터 반대 민심에 ‘수습’ 나선 오픈AI…“전력 비용 회사가 부담”Bollywood HungamaRanveer Singh’s Pralay shoot sparks BIG controversy at Mumbai’s Azad Maidan; Under-14 and Under-16 cricket matches POSTPONED; MCA FUMES over disruptionCNN TürkABD'de büyük kayıp sonrası gözler 2028'e çevrildi: İşte öne çıkan 3 isimThe South AfricanPreview of December 2026 SASSA Old-Age GrantsSoompi3 Points To Look Forward To In “What A Boss Wants”ZDF heuteAktuelle Pressemitteilungen des ZDFThe Sydney Morning HeraldBabies made a comeback - but that was before the latest rate risesVnExpress InternationalVietnam responds to other countries advising their citizens to stock up on emergency suppliesSözcüNeredeyse bedavaya ev kiralıyorlar ama istedikleri şey para değil
The Daily Newsstand · Free, Always
Friday, November 6, 2026

腾讯Hy4调用激增已十倍扩容,Agent能力开源模型第一

Translate

  11月5日,腾讯宣布因 Hy4调用激增,WorkBuddy已进行十倍扩容,以满足更多用户的体验和调用需求。同时,凭借模型能力的持续跃升,Hy4受到国际权威模型测试榜单的高度认可,在Artificial Analysis智能指数榜单上位列中国模型第二,Agent Arena榜单居开源模型第一。

  混元Hy4在WorkBuddy首发接入以来,凭借Agent能力的显著提升,受到了广大用户与开发者的积极体验和热情反馈,上线首日遇到大量用户触及额度情况。目前,腾讯已针对 Hy4推理集群进行十倍扩容,同时提升用户免费体验额度,持续动态调配资源。

  与用户侧的热烈反响相呼应,Hy4在多项海外权威评测中同步交出亮眼成绩。5日,全球权威大模型评测平台 Artificial Analysis发布最新智能指数榜单,腾讯Hy4在724款模型中跻身前列,在中国模型中位列第二,仅次于Kimi K3.1 Pro,领先于Qwen3.8 Max、DeepSeek V4.1 Flash等模型。

  Artificial Analysis被业内视为当前衡量大模型综合智力水平的重要参考之一,该指数并非单一跑分,而是一个综合了覆盖复杂推理、代码生成与调试、工具调用、多模态理解、长程规划以及科学问题求解等核心能力的高难度评测,对模型的泛化能力与工程实用性要求较高。Hy4在这一综合评价体系与 Grok4.7(xhigh)并列,显示出腾讯在基础模型能力上的扎实积累。

  Artificial Analysis也提到,在成本效率上,Hy4单任务成本约0.78美元,在同等分数区间内显著低于 Grok4.7(xhigh)和 Qwen3.8 Max,价格分别比这两个模型便宜10倍和13倍,体现出较强的能力成本平衡。

  此外,在更侧重真实智能体任务中的表现的 Agent Arena最新榜单上,腾讯 Hy4位列总榜第13名,居全球开源模型第一。凭借较高性价比,Hy4同时位于该榜单的 Pareto Frontier(帕累托前沿线)上,相比成本更高的 Claude Opus5.5等顶级模型,在显著降低成本的同时仍保持正分表现。

  11月4日,腾讯混元最新一代大语言模型 Hy4发布,模型能力的再次全面跃升,相比于 Hy4 preview,Hy4在预训练和后训练上显著扩展了数据和训练规模,首次引入视觉理解能力,并大幅提升了指令遵循、长文理解、长程智能体等综合效果。

View the original on 中国新闻网 →

KioskNews shows a cleaned-up reading view extracted from the publisher’s page — the original always lives on their site, not ours.