中国开源模型的“最大甲方”有了 腾讯WorkBuddy成开放权重模型最大客户China's Open-Source Models Find Their Biggest Customer: Tencent WorkBuddy Becomes the Largest Buyer of Open-Weight Models


AI free token/AI免费积分

中国开源模型的“最大甲方”有了 腾讯WorkBuddy成开放权重模型最大客户

本文来源:https://www.163.com/dy/article/LC7GLCHU0511I947.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

互联网大厂们正在成为开源大模型的新甲方。

腾讯总裁刘炽平在三季度业绩电话会上透露,WorkBuddy已成为中国几乎所有开放权重模型的最大客户。

近期,国内头部办公Agent密集调整产品策略。继腾讯WorkBuddy率先采用多模型路线后,字节跳动旗下豆包工作接连扩大第三方模型接入范围,阿里千问办公也重新向外部模型敞开大门。

从腾讯到字节、阿里,几家互联网大厂的办公Agent正在走向相似的产品路线。

随着这一趋势逐渐明朗,拥有庞大用户规模、能够持续承接工作任务的办公Agent有望成为上游模型厂商新的超级客户。

但当路线趋于一致后,大厂既要通过办公Agent争夺用户和工作入口,也要在不断增长的Token消耗中控制成本。

谁能汇集更多模型需求决定了业务规模;谁能更高效地调度这些模型,则开始影响商业化的成色。

WorkBuddy能够成为中国几乎所有开放权重模型最大客户,一方面与之市场影响力有关,另一方面则是因为其较早采用多模型策略。

今年3月上线时,WorkBuddy便支持接入DeepSeek、GLM、Kimi等第三方模型,此后又持续扩大模型选择范围。

这一路线 在过去一段时间里并 未成为国内办公 Agent的主流选择。

今年8月25日,字节跳动正式推出豆包工作。产品上线初期主要依靠自研豆包模型提供服务,并未向普通用户开放第三方模型选择;同期,阿里旗下的千问办公同样以自研模型为主。

但随着WorkBuddy持续扩大领先优势,其他大厂竞争压力突增。

10月29日,豆包工作率先向团队和企业订阅用户开放第三方模型。11月12日,又进一步扩大模型接入范围,在原有自研模型之外,新增Kimi K3.1 Pro、GLM 5.3、DeepSeek V4.1 Flash等第三方模型。

仅4天后,千问办公也迅速跟进,一口气上线10款模型,包括自研Qwen4-Max,以及Kimi、GLM、DeepSeek、MiniMax、Step等第三方模型。

事实上,阿里对于是否开放第三方模型曾有过摇摆。

今年8月,千问办公曾短暂引入GLM、DeepSeek等外部模型,但相关模型仅上线数日便从产品端下线,模型选择重新回到以自研模型为主的状态。

时隔三个月,千问办公再次向第三方模型敞开大门。

与此前的短暂尝试不同,这一次,千问办公不仅向个人和企业标准版用户开放模型选择,还允许企业管理员根据部门、用户组及员工配置模型使用权限。

从腾讯率先采用多模型聚合,到字节扩大第三方模型接入,再到阿里重新开放模型选择,国内几家头部大厂的办公 Agent最终走向了相似的产品路线。

从需求端来看,对办公Agent而言,用户最终需要的是一份报告、一张表格、一段代码,或者一项能够直接交付的工作。

不同任务对模型能力、响应速度和成本的要求并不相同。单一模型未必能够在所有场景中保持优势,通过接入多家模型,用户可以根据不同任务选择更合适的模型。

这种产品路线的收敛恰恰给上游模型厂商带来新的机会。

过去,智谱、月之暗面、MiniMax等独立模型厂商主要通过自身产品、开发者生态、API服务及企业客户扩大模型使用量。如今,拥有数百万乃至千万级日活跃用户的办公Agent也开始成为一类值得争取的大客户。

这种变化已经体现在部分模型厂商的新产品发布策略中。

今年10月初,WorkBuddy曾接入一款名为Space Bunny的匿名模型。直到10月19日MiniMax正式发布M3.1 Flash,外界才得知Space Bunny正是这款模型的早期版本。

据MiniMax披露,Space Bunny在匿名运行期间累计处理超过200万亿Token。

随着腾讯、字节、阿里相继走向多模型路线,各家能够获得的底层模型能力开始趋同,单纯增加模型选择已难以形成持续的差异化优势。

如何整合不同模型的能力、提高任务交付效率,并控制模型调用成本,开始成为更现实的竞争问题。

对于这些正在成为大模型重要客户的互联网大厂而言,模型调用规模不断扩大,也意味着需要承担更高的运行成本。

腾讯管理层日前在财报电话会上明确提出,未来将通过提高付费用户比例、增加每名付费用户的Token支出,以及推理优化、模型定制和算力效率提升等方式,逐步改善单位经济效益。

一方面,办公Agent需要通过更强的任务交付能力提高用户付费意愿,并推动使用量增长;另一方面则需要控制模型推理和任务执行成本,避免收入增长被持续扩大的资源消耗抵消。

不过,随着调用规模扩大,互联网大厂在承担更高成本的同时,也可能获得新的议价筹码。

庞大的采购规模也意味着有机会争取比公开API报价更低的价格。

一家已将旗下模型接入WorkBuddy的厂商人士向华尔街见闻·全天候科技确认:“在模型API的商业合作中,折扣率会根据客户的调用规模确定。”

“企业客户可以通过渠道获得低于官网公开报价的模型API服务,实际价格存在一定的协商空间。”南方一家从事大模型API代理及销售业务的企业人士向全天候科技表示。

Anthropic在官方API定价规则中明确表示,大用量客户可以根据具体情况单独协商批量折扣,企业客户还可以获得定制价格及其他商业条款。

这也可能进一步影响互联网大厂在办公Agent之间的竞争。

随着WorkBuddy、豆包工作和千问办公相继走向多模型路线,各家能够获得的底层模型供给逐渐趋同,但实际支付的采购价格未必相同。

对于拥有更大模型调用规模的平台而言,庞大的采购需求有望转化为更强的议价能力,进而获得更低的单位调用价格。

一旦这种优势形成,大厂不仅能够降低模型使用成本,也可能通过更有竞争力的产品定价和服务额度争夺用户,进一步扩大调用规模,形成正向循环。

这只是竞争的一部分。真正交付一项任务还涉及模型调用次数、Token消耗,以及执行失败后的重试成本,这也对办公Agent背后的Harness提出更高要求。

Harness作为Agent的运行框架,不仅决定如何组织这些步骤,还会影响模型调用次数、上下文管理及任务失败后的重试策略。

即使使用相同的底层模型,不同Harness也可能产生显著不同的资源消耗。

今年9月,字节Seed团队与外部研究人员发表《HarnessDev》论文,专门研究不同Harness设计对Agent任务表现及运行效率的影响。

研究人员在其中一组实验中,统一使用Gemini 3.1 Pro作为底层执行模型,让不同Harness完成相同的Terminal-Bench 2.1测试。

结果显示,由GPT-5.5构建的Harness,平均执行Token消耗约为2430万个;由Claude Opus 4.8构建的Harness,平均消耗约9430万个,相差约3.9倍。

两组实验的任务成功率分别为49.4%和52.4%。后者虽然获得了略高的成功率,但Token消耗却增加了近3倍。

这也意味着,随着未来当几家互联网大厂的办公Agent陆续接入GLM、Kimi、DeepSeek等更多第三方模型时,拥有相似的模型供给并不意味着它们能够以相同的成本完成工作。

从采购到执行,互联网大厂实际上需要算清两笔账 :

一笔是向模型厂商采购模型能力的价格。随着调用规模扩大,大厂有机会通过规模采购降低单位成本 ;

另一笔则是完成实际工作所需要的综合成本。它不仅取决于模型报价,还与模型路由、 Harness运行效率 等 密切相关。

这两种成本优势还可能相互强化,形成正向循环。

对于拥有更大调用规模的办公Agent而言,更有利的采购价格意味着相同预算能够支撑更多模型调用和产品试验。而持续积累的真实任务反馈,又有助于大厂优化模型路由、任务拆解和执行流程,加快Harness的迭代。

随着多模型路线逐渐收敛,几家互联网大厂的办公Agent能够获得的底层模型能力越来越接近,但采购成本、Harness效率与产品迭代速度仍可能拉开差距,此刻难言胜负。

*本文为全天候科技原创作品,未经授权不得转载,如需转载,请在后台回复“转载”二字,获取转载格式要求。

China's Open-Source Models Find Their Biggest Customer: Tencent WorkBuddy Becomes the Largest Buyer of Open-Weight Models

本文来源:https://www.163.com/dy/article/LC7GLCHU0511I947.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

China's big internet companies are becoming the new major clients of open-source large models.

Tencent President Liu Chiping revealed on the third-quarter earnings call that WorkBuddy(了解更多) has become the largest customer of nearly all open-weight models in China.

Recently, China's leading office agents have been intensively adjusting their product strategies. Following Tencent WorkBuddy's pioneering adoption of a multi-model route, ByteDance's Doubao Work has repeatedly expanded its third-party model integration, and Alibaba's Qwen Office has once again opened its doors to external models.

From Tencent to ByteDance and Alibaba, the office agents of several major internet companies are moving toward similar product routes.

As this trend becomes increasingly clear, office agents with huge user bases and the ability to continuously take on work tasks are poised to become new super-clients for upstream model vendors.

But once their routes converge, these major companies must both compete for users and work entry points through their office agents and control costs amid ever-growing Token consumption.

Whoever can aggregate more model demand determines the scale of the business; whoever can dispatch these models more efficiently is beginning to affect the quality of commercialization.

WorkBuddy's ability to become the largest customer of nearly all open-weight models in China is related, on one hand, to its market influence, and on the other, to its early adoption of a multi-model strategy.

When it launched in March this year, WorkBuddy already supported integration with third-party models such as DeepSeek, GLM, and Kimi, and has since continued to expand its range of model choices.

For a period in the past, this route was not the mainstream choice among domestic office agents.

On August 25 this year, ByteDance officially launched Doubao Work. In its initial period, the product mainly relied on its in-house Doubao model to provide services and did not open third-party model choices to ordinary users; during the same period, Alibaba's Qwen Office was also primarily based on its own models.

But as WorkBuddy continued to widen its lead, competitive pressure on the other major companies surged.

On October 29, Doubao Work took the lead in opening third-party models to team and enterprise subscribers. On November 12, it further expanded its model integration, adding third-party models such as Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash on top of its existing in-house models.

Just four days later, Qwen Office quickly followed suit, launching 10 models in one go, including its in-house Qwen4-Max as well as third-party models such as Kimi, GLM, DeepSeek, MiniMax, and Step.

In fact, Alibaba had once wavered on whether to open up to third-party models.

In August this year, Qwen Office briefly introduced external models such as GLM and DeepSeek, but these models were taken offline from the product after only a few days, and model selection returned to being dominated by in-house models.

Three months later, Qwen Office once again opened its doors to third-party models.

Different from the earlier brief attempt, this time Qwen Office not only opened model selection to individual and enterprise standard-edition users, but also allowed enterprise administrators to configure model usage permissions based on departments, user groups, and employees.

From Tencent taking the lead in multi-model aggregation, to ByteDance expanding third-party model integration, to Alibaba reopening model selection, the office agents of several of China's top major companies ultimately converged on similar product routes.

From the demand side, for office agents, what users ultimately need is a report, a spreadsheet, a piece of code, or a task that can be delivered directly.

Different tasks have different requirements for model capability, response speed, and cost. A single model may not maintain an advantage in all scenarios; by integrating multiple models, users can select the more suitable model for different tasks.

This convergence of product routes is precisely bringing new opportunities to upstream model vendors.

In the past, independent model vendors such as Zhipu, Moonshot AI, and MiniMax expanded model usage mainly through their own products, developer ecosystems, API services, and enterprise customers. Today, office agents with millions or even tens of millions of daily active users are also becoming a category of major clients worth pursuing.

This shift is already reflected in the new product launch strategies of some model vendors.

In early October this year, WorkBuddy integrated an anonymous model called Space Bunny. It was not until October 19, when MiniMax officially released M3.1 Flash, that outsiders learned Space Bunny was an early version of this model.

According to MiniMax, Space Bunny processed a cumulative total of over 200 trillion Tokens during its anonymous run.

As Tencent, ByteDance, and Alibaba successively moved toward multi-model routes, the underlying model capabilities available to each began to converge, and simply adding more model choices could hardly form a sustained differentiated advantage.

How to integrate the capabilities of different models, improve task delivery efficiency, and control model invocation costs is becoming a more practical competitive issue.

For these major internet companies that are becoming important clients of large models, the ever-expanding scale of model invocation also means bearing higher operating costs.

Tencent's management recently stated clearly on an earnings call that in the future it will gradually improve unit economics by increasing the proportion of paying users, raising Token spending per paying user, and through inference optimization, model customization, and improvements in computing power efficiency.

On one hand, office agents need to increase users' willingness to pay and drive usage growth through stronger task delivery capabilities; on the other hand, they need to control model inference and task execution costs, so that revenue growth is not offset by continuously expanding resource consumption.

However, as invocation scale expands, major internet companies, while bearing higher costs, may also gain new bargaining chips.

A huge procurement scale also means an opportunity to negotiate prices lower than published API quotes.

A representative of a vendor that has integrated its models into WorkBuddy confirmed to Wallstreetcn · AllTime Tech: 'In commercial cooperation on model APIs, the discount rate is determined according to the customer's invocation scale.'

'Enterprise customers can obtain model API services through channels at prices lower than the official website's public quotes; there is some room for negotiation in the actual price,' a representative of a southern China company engaged in large-model API agency and sales told AllTime Tech.

Anthropic explicitly states in its official API pricing rules that high-volume customers can individually negotiate bulk discounts depending on the specific situation, and enterprise customers can also receive custom pricing and other commercial terms.

This may also further affect competition among major internet companies in the office agent space.

As WorkBuddy, Doubao Work, and Qwen Office successively moved toward multi-model routes, the underlying model supply available to each gradually converged, but the procurement prices actually paid may not be the same.

For platforms with larger model invocation scales, huge procurement demand is expected to translate into stronger bargaining power, and thus lower per-invocation prices.

Once such an advantage takes shape, major companies can not only lower model usage costs, but also compete for users through more competitive product pricing and service quotas, further expanding invocation scale and forming a positive feedback loop.

This is only part of the competition. Actually delivering a task also involves the number of model invocations, Token consumption, and the cost of retries after execution failures, which places higher demands on the Harness behind the office agent.

As the operating framework of an agent, the Harness not only determines how these steps are organized, but also affects the number of model invocations, context management, and retry strategies after task failures.

Even when using the same underlying model, different Harnesses can produce significantly different resource consumption.

In September this year, ByteDance's Seed team and external researchers published a paper titled 'HarnessDev,' specifically studying the impact of different Harness designs on agent task performance and operational efficiency.

In one set of experiments, the researchers uniformly used Gemini 3.1 Pro as the underlying execution model, having different Harnesses complete the same Terminal-Bench 2.1 test.

The results showed that the Harness built with GPT-5.5 consumed an average of about 24.3 million Tokens in execution, while the Harness built with Claude Opus 4.8 consumed about 94.3 million on average—a difference of about 3.9 times.

The task success rates of the two experiments were 49.4% and 52.4%, respectively. Although the latter achieved a slightly higher success rate, its Token consumption increased by nearly threefold.

This also means that, in the future, as the office agents of several major internet companies successively integrate more third-party models such as GLM, Kimi, and DeepSeek, having similar model supplies does not mean they can complete work at the same cost.

From procurement to execution, major internet companies actually need to carefully calculate two accounts:

One is the price of purchasing model capabilities from model vendors. As invocation scale expands, major companies have the opportunity to lower unit costs through bulk procurement;

The other is the overall cost required to complete the actual work. It depends not only on model quotes, but is also closely related to model routing, Harness operating efficiency, and other factors.

These two cost advantages can also reinforce each other, forming a positive feedback loop.

For office agents with larger invocation scales, more favorable procurement prices mean the same budget can support more model invocations and product experiments. Meanwhile, the continuous accumulation of real task feedback helps major companies optimize model routing, task decomposition, and execution workflows, accelerating Harness iteration.

As multi-model routes gradually converge, the underlying model capabilities available to the office agents of several major internet companies are becoming increasingly similar, but gaps may still open up in procurement costs, Harness efficiency, and product iteration speed—it is too early to tell who will win.

*This article is an original work of AllTime Tech and may not be reproduced without authorization. If you wish to reproduce it, please reply with the word 'reprint' (转载) in the backend to obtain the reproduction format requirements.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry