广告
加载中

姚顺雨到岗5个月 腾讯混元交出第一份答卷

亿邦动力 2026-04-23 18:11
亿邦动力 2026/04/23 18:11

邦小白快读

EN
全文速览

腾讯混元团队发布Hy3 preview语言模型并开源,带来了显著的性能提升和实用成本优化。

1. 模型采用混合专家架构,总参数295B,激活参数仅21B,支持长达256K上下文,在复杂推理、代码生成和智能体能力上大幅提升。

2. 推理效率提高40%,输入价格最低降至每百万Tokens 1.2元,并推出最低28元/月的个人套餐,大大降低使用门槛。

3. 在内部900多个业务场景测试中,模型提升产品体验如元宝端意图理解精度、QQ小Q助手文本创作质量,以及CodeBuddy首token延迟降低54%、成功率超99.99%。

腾讯通过Hy3 preview模型强化品牌营销策略并推动产品研发创新,反映出消费趋势中的AI应用深化。

1. 品牌营销上,坚持开源与阿里等转向闭源形成对比,抢夺AI开放生态声量,吸引开发者社区反馈。

2. 产品研发中,模型深度协同于元宝端、腾讯文档AI PPT和QQ小Q助手,提升意图理解精度20%、生成成功率20%,体现用户行为对高效创作的需求。

3. 消费趋势显示AI在游戏如和平精英AI NPC和文档生成中带来可量化体验优化,价格竞争策略(输入低至1.2元/百万Tokens)迎合中小企业低成本需求。

Hy3 preview的开源和低价策略提供市场增长机会与风险管理启示,凸显可学习商业模式。

1. 政策解读上,国内大模型开源生态转向中,腾讯坚持开源带来合作机会,如免费流量吸引开发者反馈。

2. 增长市场机会在于低价套餐(个人28元/月)满足消费需求变化,服务中小企业AI应用。

3. 风险提示包括竞争环境中的性价比挑战,但可学习点如姚顺雨的实用性原则(能力体系化、评测真实性)帮助控制成本,提升正面影响如内部测试中495步复杂Agent工作流的稳定性。

Hy3 preview模型为制造业提供产品设计启示和数字化商业机会,推进电商智能化。

1. 产品生产和设计需求上,模型的高效参数(295B总/21B激活)和长上下文支持可用于智能设计自动化。

2. 商业机会源于低价AI服务(输入1.2元/百万Tokens),降低工厂数字转型成本,推进电商应用如内部测试中的CodeBuddy效率提升。

3. 推进数字化启示来自900多个业务场景案例,如在WorkBuddy中降低延迟、提升成功率,启发工厂优化生产流程。

Hy3 preview引入新技术解决行业痛点,展现行业发展趋势和服务解决方案。

1. 行业发展趋势显示国内大模型开源生态转向,腾讯坚持开源推动AI开放竞争。

2. 新技术如混合专家模型提升推理效率40%,并在代码和智能体方向取得SWE-Bench等评测领先,解决客户高成本痛点。

3. 解决方案包括输入价格最低1.2元/百万Tokens,个人套餐28元/月,提供可靠服务如内部测试中99.99%以上成功率,优化企业应用场景。

腾讯云TokenHub平台通过Hy3 preview服务优化商业需求与运营管理,规避风险。

1. 商业对平台的需求如长上下文256K支持和高可靠性模型(内部测试成功率99.99%),吸引企业用户。

2. 平台最新做法包括低价输入套餐(最低1.2元/百万Tokens)和开源策略,促进招商和开发者合作。

3. 运营管理中,内部测试提升产品如ima知识库问答的信息覆盖度,规避风险通过稳定驱动495步复杂工作流,确保服务连续性。

Hy3 preview反映产业新动向和商业模式创新,提出实用性问题与政策启示。

1. 产业新动向是国内开源生态微妙转向(阿里闭源、腾讯坚持开源),追求性价比原则挑战万亿参数竞赛。

2. 新问题是如何在资源有限下实现能力体系化和评测真实性(如清华数学考题领先),推动模型实用性。

3. 商业模式包括低价服务(输入1.2元/百万Tokens)和开源反馈机制,为政策法规建议如促进AI社区协作提供启示,未来方向是扩大预训练规模探索特色能力。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Tencent's Hunyuan team has released and open-sourced the Hy3 preview language model, delivering significant performance gains and practical cost optimization.

1. The model uses a Mixture of Experts architecture with 295B total parameters and only 21B activated parameters, supporting a context length of up to 256K tokens. It shows major improvements in complex reasoning, code generation, and agent capabilities.

2. Inference efficiency has increased by 40%, with input prices dropping to as low as ¥1.2 per million tokens. A personal plan starting at ¥28 per month significantly lowers the barrier to entry.

3. In internal testing across over 900 business scenarios, the model enhanced product experiences, such as improving intent understanding accuracy in Yuanbao, boosting text creation quality for QQ's Xiao Q assistant, and reducing CodeBuddy's first-token latency by 54% while achieving a success rate exceeding 99.99%.

Tencent is strengthening brand marketing strategies and driving product R&D innovation with the Hy3 preview model, reflecting the deepening application of AI in consumer trends.

1. In brand marketing, Tencent's commitment to open-source contrasts with Alibaba's shift to closed-source, aiming to capture mindshare in the open AI ecosystem and attract developer community feedback.

2. For product R&D, the model is deeply integrated into Yuanbao, Tencent Docs' AI PPT, and QQ's Xiao Q assistant, improving intent understanding accuracy by 20% and generation success rate by 20%, meeting user demand for efficient creation.

3. Consumer trends show AI delivers quantifiable experience optimizations in areas like game NPCs (e.g., Peacekeeper Elite) and document generation. The competitive pricing strategy (input as low as ¥1.2/million tokens) caters to the low-cost needs of SMEs.

Hy3 preview's open-source and low-price strategy presents market growth opportunities and risk management insights, highlighting a learnable business model.

1. Policy-wise, amidst a subtle shift in China's large model open-source ecosystem, Tencent's open-source stance creates collaboration opportunities, such as using free traffic to attract developer feedback.

2. Growth opportunities lie in affordable plans (e.g., personal plan at ¥28/month) meeting evolving consumer demand and serving SME AI applications.

3. Risks include value-for-money challenges in a competitive landscape, but learnable aspects like Yao Shunyu's practicality principle (systematizing capabilities, ensuring evaluation authenticity) help control costs and enhance positive impacts, such as the stability of 495-step complex Agent workflows in internal tests.

The Hy3 preview model offers insights for product design and digital commercial opportunities in manufacturing, advancing e-commerce intelligence.

1. For product production and design needs, the model's efficient parameters (295B total/21B activated) and long context support can be applied to intelligent design automation.

2. Commercial opportunities arise from low-cost AI services (input at ¥1.2/million tokens), reducing digital transformation costs for factories and promoting e-commerce applications, as seen in CodeBuddy's efficiency gains during internal testing.

3. Digital advancement insights come from over 900 business use cases, such as reduced latency and improved success rates in WorkBuddy, inspiring factories to optimize production processes.

Hy3 preview introduces new technologies addressing industry pain points, showcasing development trends and service solutions.

1. Industry trends indicate a shift in China's large model open-source ecosystem, with Tencent's open-source approach promoting open AI competition.

2. New technologies like the MoE architecture improve inference efficiency by 40% and achieve leading scores on benchmarks like SWE-Bench for code and agent capabilities, solving the high-cost pain point for clients.

3. Solutions include input prices as low as ¥1.2/million tokens and a personal plan for ¥28/month, providing reliable services with over 99.99% success rates in internal tests, optimizing enterprise application scenarios.

Tencent Cloud's TokenHub platform optimizes commercial needs and operational management with Hy3 preview services, while mitigating risks.

1. Commercial platform demands, such as 256K context support and high-reliability models (99.99% success rate in internal tests), attract enterprise users.

2. The platform's latest initiatives include low-cost input packages (minimum ¥1.2/million tokens) and an open-source strategy to foster merchant acquisition and developer collaboration.

3. In operational management, internal testing enhanced products like the ima knowledge base Q&A's information coverage. Risks are mitigated by stably driving 495-step complex workflows, ensuring service continuity.

Hy3 preview reflects new industry trends and business model innovations, raising practical questions and policy implications.

1. A key industry trend is the subtle shift in China's open-source ecosystem (Alibaba closing source vs. Tencent staying open), challenging the trillion-parameter race with a value-for-money principle.

2. A new research question is how to systematize capabilities and ensure evaluation authenticity (e.g., leading on Tsinghua math problems) with limited resources to advance model practicality.

3. The business model includes low-cost services (input at ¥1.2/million tokens) and an open-source feedback mechanism, offering insights for policy recommendations like promoting AI community collaboration. Future directions involve scaling pre-training to explore distinctive capabilities.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

【亿邦原创】4月23日,腾讯混元团队发布了Hy3 preview语言模型并同步开源。这是2026年2月腾讯重建预训练和强化学习基础设施之后,混元训练的第一个新模型。

Hy3 preview是一款快慢思考融合的混合专家模型,总参数295B,激活参数仅21B,最大支持256K上下文长度。在复杂推理、指令遵循、代码、智能体等能力上,Hy3 preview均实现了大幅提升,而推理效率提升了40%,输入价格最低降至每百万Tokens 1.2元。

这是姚顺雨归国加盟腾讯半年多后交出的一份阶段性成果。

2025年底,腾讯调整架构,任命前OpenAI资深研究员姚顺雨为首席AI科学家,领衔新成立的AI Infra部、AI Data部及大语言模型部。马化腾曾将这一系列举措定义为对混元的“深度重构”——从外在的人才引入、团队革新,到内在的基础设施协同、产品协同,都发生改变。2026年2月,姚顺雨带队完成了预训练和强化学习基础设施的重建,Hy3 preview便是这一重建的成果。

姚顺雨为混元定下的三大实用性原则——“能力体系化、评测真实性、追求性价比”,核心目标就是在资源有限的条件下,为腾讯“赢回”竞争力。

在同行都卷万亿参数时,Hy3预览版选了一个总参数295B、激活仅21B的规格。在能力表现上,Hy3 preview在清华数学博资考、全国中学生生物学联赛等真实考题中取得了国内领先的分数。代码和智能体方向提升最为显著,在SWE-Bench Verified、Terminal-Bench 2.0以及ClawEval等主流评测中均取得有竞争力的结果。

更关键的是成本大幅下降,在腾讯云TokenHub上,输入价格最低1.2元/百万Tokens,并推出了最低28元/月的个人套餐。

正式上线前,Hy3 preview也在腾讯内部超过900个业务场景中完成了产品测试。在元宝端,模型与产品深度协同,提升了意图理解精度和文本创作质量;在ima知识库问答中,长文处理和信息覆盖度表现突出;在CodeBuddy和WorkBuddy上,首token延迟降低54%,成功率提升至99.99%以上,已稳定驱动最长495步的复杂Agent工作流。此外,在和平精英AI NPC、腾讯文档AI PPT、QQ小Q助手等场景中,Hy3 preview均带来了可量化的体验提升——例如AI PPT生成成功率提升20%,耗时缩短20%。

姚顺雨将之定义为“混元大模型重建的第一步”,并表示开源是为了获得真实反馈,帮助提升正式版的实用性。

近期国内大模型开源生态正经历一场微妙转向。阿里Qwen系列已全面转向闭源,仅提供API调用;MiniMax、智谱的最新模型也纷纷闭源。主流玩家中,坚持开源的只剩下DeepSeek与腾讯混元。

Hy3 preview通过开源的免费流量和社区反馈,不仅能为正式版快速迭代铺路,也能为腾讯在AI开放生态中抢夺有利声量。

当然,Hy3 preview只是混元大模型重建的第一步,为“资源相对有限”的腾讯在AI下半场找回自己的节奏和打法,未来还会继续扩大预训练规模、探索特色模型能力。

文章来源:亿邦动力

广告
微信
朋友圈

这么好看,分享一下?

朋友圈 分享

APP内打开

+1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0