广告
加载中

Hy3正式发布 腾讯的AI焦虑缓解了吗?

胡镤心 2026-07-06 16:20
胡镤心 2026/07/06 16:20

邦小白快读

EN
全文速览

本次腾讯正式发布混元Hy3大模型,核心重点信息和可供普通用户、开发者实操的干货整理如下:

1. 核心性能与落地成果:Hy3总参数295B、激活参数21B,属于中等体量实战型模型,实测性能较前代Hy2提升超40%,内部专家盲测表现优于GLM5.1,Token消耗量曾位列全球周榜第二,已经接入腾讯多个核心产品,给产品性能和用户留存带来明显提升。

2. 可获取的实操资源:Hy3采用商业友好度高的Apache2.0开源协议,全球开发者都可以下载并免费商用,目前已经陆续上线多个国内外开发平台,同时首日接入Huggingface、魔搭两大开源社区,获取和使用都非常方便。

3. 定价亲民门槛低:Hy3定价普惠,输入1元/百万tokens、输出4元/百万tokens,命中缓存仅0.25元/百万tokens,个人开发者和中小团队都负担得起。

针对品牌商关注的产品研发、定价竞争、消费趋势等内容,整理核心干货如下:

1. 可借鉴的产品研发模式:Hy3是从腾讯真实业务场景中打磨出来的模型,研发过程中在50多个业务收集用户反馈,和内部业务线深度协同设计,提前在真实场景验证迭代,这种贴近需求的研发模式值得各领域品牌借鉴。

2. 差异化定价与竞争策略:Hy3避开头部厂商的超大参数竞赛,走中等体量、实用普惠的差异化路线,定价远低于行业常规水平,通过开源开放快速抢占开发者市场,这种错位竞争策略值得品牌参考。

3. 未来消费趋势机会:接下来Hy3将落地微信14亿月活生态,AI智能体将成为新的用户交互入口,消费品牌可以提前布局微信AI生态的营销和获客机会,抢占新流量红利。

针对卖家关注的增长机会、风险提示、合作机会等内容,核心干货整理如下:

1. 降低AI应用开发门槛:Hy3采用免费商用的开源协议,定价普惠,大大降低了卖家开发自有AI工具的成本,中小卖家可以基于该模型开发专属客服、选品分析、内容制作等运营工具,不需要投入高额成本训练大模型。

2. 接入方便门槛低:Hy3发布首日就接入了全球主流开源社区和多个海外开发平台,卖家和开发者可以快速获取接入,不需要复杂的部署流程,使用成本很低。

3. 未来增长机会提示:接下来基于Hy3的微信AI智能体已经进入灰度测试,微信14亿月活将诞生新的AI流量生态,卖家可以提前关注该领域的开放政策,提前布局新的流量增长机会,抢占先发优势。

针对工厂关注的数字化转型、产品设计需求、商业机会等内容,整理核心干货如下:

1. AI赋能数字化转型的新选择:Hy3已经在设计、办公、软件开发等多个场景验证了能力,推理效率提升40%以上,幻觉率仅5.4%,单意图识别准确率超过95%,工厂可以依托该模型快速开发适合自身的产品设计、生产管理、客户服务等数字化工具,不需要从零训练大模型,降低转型成本。

2. 数字化转型的启示:工厂推进数字化不需要盲目追求超大参数模型,中等体量适配真实场景的模型性价比更高、落地更快,Hy3从业务场景中来的打磨模式值得工厂借鉴。

3. 新商业机会:接下来Hy3将落地微信AI生态,面向C端的消费类工厂,可以提前关注AI智能体带来的新用户交互、获客模式,探索直接触达C端用户的新路径。

针对AI服务商关注的行业趋势、客户痛点、解决方案等内容,核心干货整理如下:

1. 当前行业发展趋势:大模型行业已经从早期的参数竞赛转向真实场景实战能力比拼,客户不再单纯追求总参数大小,更看重模型的任务完成效率、推理速度和准确率,中等体量高适配性的模型更受市场欢迎,市场空间广阔。

2. 企业客户的核心痛点:很多企业布局AI过程中,存在多模型并行、底座不统一的问题,导致开发者需要在不同平台切换,效率低下,算力浪费,搭建统一的内部AI基座已经成为企业客户的核心需求。

3. 可借鉴的业务发展模式:Hy3采用开源开放+普惠定价的模式快速拓展市场,通过提前在大量真实场景打磨产品提升能力,同时多渠道同步上线覆盖全球开发者,这种获客和产品打磨模式值得To B AI服务商借鉴。

针对平台商关注的商家需求、招商运营、风向规避等内容,核心干货整理如下:

1. 模型方对平台的核心需求:大模型方发布新产品后,需要快速接入主流开源社区和第三方开发者平台,触达全球开发者,扩大产品影响力,首日上线多平台已经成为头部开源大模型的常规操作。

2. 平台运营和招商的启示:开源社区和开发者平台可以通过优先引入优质实用型开源大模型,吸引更多开发者入驻,提升平台活跃度和用户粘性,Hy3首日接入Huggingface和魔搭,就给两个平台带来了大量流量。

3. 行业风向提示:当前高性价比的中等体量实战大模型更受开发者欢迎,调用量增长迅速,平台可以针对性调整招商方向,引入这类模型满足市场需求,同时要提前布局海外渠道,适配全球开发者的使用需求,提升平台竞争力。

针对研究者关注的产业新动向、商业模式、企业战略等内容,核心干货整理如下:

1. 大模型产业竞争新动向:当前大模型产业竞争格局已经发生变化,头部厂商不再一味追求超大参数,差异化竞争路线开始出现,腾讯Hy3走中等体量、普惠实战的差异化路线,错位竞争头部超大参数模型,已经取得不错的市场成果,Token消耗量位列全球第二,证明该路线的可行性。

2. 互联网大厂AI战略新变化:腾讯此前AI布局分散,内部多模型并行导致效率低下,本次通过Hy3统一内部AI能力基座,解决了算力浪费和效率问题,缓解了腾讯的AI焦虑,为后续微信生态AI落地打好基础,这反映了大厂AI布局从分散到整合的新趋势。

3. 大模型商业化新商业模式:开源开放+低定价普惠已经成为大模型厂商快速获取市场份额、打磨产品的新型商业模式,商业友好的开源协议更能吸引开发者生态,这种模式值得深入研究。

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 has officially launched its Hunyuan Hy3 large language model. Below we summarize the key updates and actionable takeaways for general users and developers:

1. Core performance and production deployments: Hy3 is a mid-sized production-grade model with 295B total parameters and 21B active parameters. Independent testing shows its performance improves by over 40% compared to its predecessor Hy2, and it outperforms GLM5.1 in blind expert evaluations. It ranked second globally in weekly token consumption, and is already integrated into multiple core Tencent products, delivering measurable improvements to product performance and user retention.

2. Accessible developer resources: Hy3 is released under the business-friendly Apache 2.0 open-source license, allowing free commercial use and download for developers worldwide. It is already available on multiple major domestic and international development platforms, and launched on two leading open-source communities, Hugging Face and ModelScope, on day one for easy access and adoption.

3. Affordable, low-threshold pricing: Hy3 is priced for mass accessibility: 1 RMB per million input tokens, 4 RMB per million output tokens, and just 0.25 RMB per million cached tokens. This pricing fits the budgets of independent developers and small-to-medium teams.

Below we summarize key takeaways for brands, focused on product development, competitive pricing and consumer trends:

1. A replicable product development approach: Hy3 was built and refined specifically for Tencent’s real business scenarios. Tencent collected user feedback across over 50 business lines during development, collaborated closely with internal teams, and iterated on the model through early real-world testing. This demand-centric development framework is a valuable reference for brands across all sectors.

2. A actionable differentiated competitive strategy: Instead of competing with top players in a race for ever-larger parameter counts, Hy3 adopts a differentiated positioning as a mid-sized, practical, and affordable model. It is priced far below the industry standard, and is using open-sourcing to quickly capture developer mindshare. This off-grid competitive strategy is a useful blueprint for brands.

3. Upcoming consumer trend opportunities: Hy3 will soon be deployed in WeChat’s 1.4-billion monthly active user ecosystem, where AI agents will emerge as a new user interaction entry point. Consumer brands can get a head start on marketing and customer acquisition in WeChat’s AI ecosystem to capture early access to this new traffic dividend.

Below we summarize key takeaways for sellers, focused on growth opportunities, risk notes, and partnership openings:

1. Lower barriers to building AI tools: Released under a free commercial open-source license with affordable pricing, Hy3 drastically cuts the cost for sellers to build custom AI tools. Small and medium-sized sellers can build dedicated tools for customer service, product selection analysis, and content creation, without bearing the high cost of training a large foundation model from scratch.

2. Easy integration with low barriers: Hy3 was available on all major global open-source communities and multiple international development platforms on launch day. Sellers and developers can access and integrate the model quickly, with no complex deployment process and very low usage costs.

3. Upcoming growth opportunities: WeChat AI agents built on Hy3 are already in灰度 testing. The 1.4-billion MAUs of WeChat will spawn a new AI-powered traffic ecosystem. Sellers can monitor upcoming opening policies for this space, prepare early to tap into the new growth channel, and secure first-mover advantage.

Below we summarize key takeaways for factories, focused on digital transformation, product design needs, and new business opportunities:

1. A new option for AI-powered digital transformation: Hy3 has already validated its performance across design, office work, software development and other use cases, delivering over 40% higher inference efficiency, a 5.4% hallucination rate, and over 95% single-intent recognition accuracy. Factories can quickly build custom digital tools for product design, production management, customer service and more on top of Hy3, eliminating the need to train a large model from scratch and cutting transformation costs.

2. Key insights for digital transformation: Factories do not need to blindly chase extremely large parameter models for digital transformation. Mid-sized models tailored to real use cases deliver better cost efficiency and faster deployment. Hy3’s approach of refining the model directly through business end scenarios is a useful reference for factories.

3. New business opportunities: Hy3 will soon be integrated into the WeChat AI ecosystem. Consumer-facing factories can get an early look at new user interaction and customer acquisition models enabled by AI agents, and explore new paths to directly reach end consumers.

Below we summarize key takeaways for AI service providers, focused on industry trends, customer pain points, and solution frameworks:

1. Current industry trends: The large model industry has shifted from the early era of parameter racing to competition based on real-world performance. Clients no longer prioritize raw total parameter size, and instead value task efficiency, inference speed, and accuracy. Mid-sized, high-adaptability models are now in higher market demand with substantial market opportunity.

2. Core pain points for enterprise clients: Many organizations face the problem of scattered multiple models with an inconsistent base layer when adopting AI, forcing developers to switch between platforms, lowering productivity and wasting computing power. Building a unified internal AI base has become a core demand for enterprise clients.

3. A replicable business development model: Hy3 uses an open-source + affordable pricing model to quickly scale market reach, refining its capabilities through early testing across a large number of real-world scenarios, and launching simultaneously across multiple channels to reach global developers. This approach to customer acquisition and product refinement is a valuable reference for B2B AI service providers.

Below we summarize key takeaways for platform operators, focused on merchant demand, recruitment and operations, and risk mitigation:

1. Core platform needs from model developers: When launching new models, large model developers need fast integration into leading open-source communities and third-party developer platforms to reach global developers and expand product influence. Multi-platform launch on day one has become standard practice for top open-source large models.

2. Key insights for platform operations and developer recruitment: Open-source communities and developer platforms can attract more developers and boost platform activity and user retention by prioritizing high-quality, practical open-source large models. Hy3’s day-one launch on Hugging Face and ModelScope already drove substantial traffic to both platforms.

3. Industry trend guidance: Mid-sized, cost-effective production-grade large models are currently the most popular option among developers, with call volume growing rapidly. Platforms can adjust their recruitment strategy to prioritize this type of model to meet market demand. They should also build out international channels early to meet the needs of global developers and improve platform competitiveness.

Below we summarize key takeaways for researchers, focused on new industry dynamics, business models, and corporate strategy:

1. New competitive dynamics in the large model industry: The competitive landscape of the large model industry has shifted. Top industry players are no longer solely focused on pushing larger parameter counts, and differentiated competition strategies are emerging. Tencent’s Hy3 pursues a differentiated positioning as a mid-sized, affordable, production-ready model to compete off-grid against top players’ ultra-large models, and has already delivered strong market results, including the second-highest global token consumption, proving the viability of this strategic path.

2. New shifts in Big Tech AI strategy: Tencent’s previous AI布局 was fragmented, with multiple parallel internal models dragging down efficiency. With Hy3, the company has unified its internal AI capability base, resolving issues of wasted computing power and low efficiency, easing Tencent’s "AI anxiety" and laying the groundwork for future AI deployment across the WeChat ecosystem. This reflects a broader industry shift from fragmented to consolidated AI布局 among large technology companies.

3. A new commercialization model for large models: Open-sourcing paired with low, accessible pricing has emerged as a new business model for large model developers to quickly capture market share and refine products. Business-friendly open-source licenses are particularly effective at growing developer ecosystems, making this model a high-priority topic for further research.

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】7月6日,腾讯混元Hy3正式发布。从2月重建基础设施,到4月推出预览版,再到7月正式落地,腾讯用不到半年时间,走完了从底层基建到旗舰模型的完整链路。

Hy3是一款从腾讯真实业务场景中长出来的解决问题的模型。从preview到正式版,混元团队在50多个业务中收集了反馈,修复了体验问题,提升了后训练的质量和规模。

据介绍,Hy3在腾讯内部各业务线中进行深度Co-Design,预览阶段就被大规模投入到元宝、ima、CodeBuddy、WorkBuddy等腾讯核心产品中进行真实场景的验证和历练。在CodeBuddy和WorkBuddy等智能体应用中,Hy3 preview的首次响应速度提升了54%,任务平均完成时间缩短了47%,任务成功率保持在99.99%;在元宝,约80%的用户已经用上了Hy3驱动的服务,且留存率有明显提升。

在当下的旗舰模型格局中,Hy3属于中等体量、实战能打的类型。

Hy3的总参数为295B、激活参数21B,采用MoE架构,原生支持256K上下文。这个体量在旗舰模型中属于“中等偏小”。DeepSeek-V4 Pro总参数1.6T、激活49B;GLM-5.1、Kimi K2.6等也都是更大体量的选手。

在实际评测中,Hy3在SWE-Bench Verified上,Hy3 preview得分74.4%,相比Hy2的53.0%提升超过40%。在内部组织的270位专家基于真实工作的模型盲测中,Hy3(均分2.67/4)展现出优于GLM5.1(均分2.51/4)的表现,尤其在前端、数据与存储、CI/CD等类别优势显著。

在调用量上,Hy3 preview曾以3.03T的Token消耗量位列全球大模型Token消耗周榜第二,仅次于DeepSeek V4 Flash的3.11T。

对腾讯来说,Hy3给了腾讯内部一个更统一的AI能力基座。

以前腾讯做AI,总被说“慢半拍”,元宝里用DeepSeek,WorkBuddy里用混元,内部好几套模型在跑,开发者在不同平台间跳来跳去。Hy3出来后,已经在WorkBuddy/CodeBuddy、元宝、Marvis、ima等多个业务接入,在软件开发、办公生产、金融建模、前端设计、游戏制作等生产力任务上的进步尤其显著,让腾讯内部各条业务线跑在同一个底座上。

Hy3也延续了实用、普惠的模型定位,定价为输入1元/百万tokens,输出4元/百万tokens,输入命中缓存价格0.25元/百万tokens。

开源方面,Hy3采用商业友好度高的Apache2.0开源协议,全球开发者均可下载和免费商用。为进一步方便全球开发者使用,Hy3将陆续在多个海外平台上线,覆盖OpenRouter、Hermes、Kilo、Cline、OpenClaw、OpenCode、CherryStudio等,且同时“day 0”接入开源模型社区Huggingface、Modelscope魔搭平台。

Hy3能不能彻底缓解腾讯的AI焦虑,还不好说,但它至少解决了一个具体问题,腾讯终于不用再到处“借”模型了。接下来要看的,是各条业务线能不能真正用好它,尤其是能不能在微信里跑起来。

微信AI智能体“小微”已在灰度测试,有投行认为,微信AI Agent已成为“一个分阶段推出、里程碑清晰可见的项目”。而Hy3已经把推理效率提升了40%以上,幻觉率降到了5.4%,单意图识别准确率超过95%——这些数据在微信14亿月活用户面前,能跑出什么结果,才是值得看的。

Hy3已经证明了它能干活,接下来要期待它能帮微信长出什么新东西。

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

腾讯混元Hy3是什么?

腾讯混元Hy3是腾讯2024年7月正式发布的实战型大模型,总参数295B、激活参数21B,采用MoE架构,原生支持256K上下文,定价为输入1元/百万tokens,采用Apache2.0开源协议支持免费商用。

腾讯混元Hy3的性能表现怎么样?

Hy3预览版在智能体应用中首次响应速度提升54%,任务平均完成时间缩短47%,成功率达99.99%;SWE-Bench Verified得分74.4%,较Hy2提升超40%,专家盲测表现优于GLM5.1,Token消耗量曾位列全球大模型周榜第二。

腾讯混元Hy3对腾讯内部业务有什么价值?

Hy3为腾讯提供了统一的AI能力基座,解决了此前内部多套模型并行、开发者需跨平台切换的问题,目前已接入元宝、WorkBuddy等多个核心业务,在软件开发、办公生产、游戏制作等生产力任务上进步显著。

开发者可以免费使用腾讯混元Hy3吗?

可以,腾讯混元Hy3采用商业友好度高的Apache2.0开源协议,全球开发者均可下载和免费商用,目前已接入Huggingface、Modelscope魔搭平台,还将陆续在OpenRouter等多个海外平台上线。

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