Consensus on Open Source

Author: Lian Ran

Recently, the two most关注ed AI companies in China have quietly come together.

On the evening of March 11, Manus officially announced a strategic cooperation with the Alibaba Tongyi Qianwen team. In the announcement, Manus stated that both parties will provide the Tongyi Qianwen series of open-source models to realize all of Manus's functions on domestic models and computing power platforms.

According to Geek Park, the contact time between the two parties has not been long, but they quickly reached a cooperation intention after their initial contact. In fact, just last Friday, Geek Park had recommended the Manus team to connect and communicate with relevant personnel from Alibaba Group—based on the judgment that Alibaba Cloud could provide important support for the Manus startup team. In fact, Alibaba has always placed great importance on innovative projects, "They are also aware that their capabilities can help entrepreneurs."

Alibaba's actions this time can be described as very swift. From the subsequent shares in the Manus team's social circle, it seems that Alibaba Group CEO Wu Yongming may have also participated. Now, both parties have already begun to have clear next steps. This incident also reflects Alibaba's long-standing open-mindedness and the significant strength it has shown in the open-source field.

Timely Cooperation

After Manus went live, in addition to the amazement at its ability to independently solve complex tasks, there was widespread speculation about the foundational models it used.

Previously, Manus co-founder Ji Yichao had revealed on social media that Manus products utilized Claude and various fine-tuned models based on Qwen.

Image source: X

The Tongyi Qianwen reasoning model QwQ-32B is indeed one of the most关注ed AI large models during this period.

On the same day Manus was released, the Tongyi Qianwen reasoning model QwQ-32B was also released and open-sourced. An official report from Alibaba Technology stated that through large-scale reinforcement learning, Qianwen QwQ-32B has achieved a qualitative leap in mathematics, coding, and general capabilities, with overall performance comparable to DeepSeek-R1:

In a series of authoritative benchmark tests, the Qianwen QwQ-32B model performed excellently, almost completely surpassing OpenAI-o1-mini, and is on par with the strongest open-source reasoning model DeepSeek-R1: In the AIME24 evaluation set testing mathematical abilities and the LiveCodeBench assessing coding abilities, Qianwen QwQ-32B performed comparably to DeepSeek-R1, far surpassing o1-mini and the same-sized R1 distilled model; in the "Most Difficult LLMs Evaluation List" LiveBench led by Meta's chief scientist Yang Likun, Google's instruction-following ability IFEval evaluation set, and the BFCL test assessing accurate function or tool invocation proposed by the University of California, Berkeley, Qianwen QwQ-32B's scores exceeded those of DeepSeek-R1.

Previously, it was reported that Manus incurs a cost of $2 for each invocation, breaking down and solving a complex task could consume millions of tokens. As the number of users increases, task queue delays have begun to appear—even under a system where only a few people can use it through invitation codes.

Manus's multi-agent architecture and full-link autonomous execution capability mean it needs to handle more complex tasks and more data, which naturally increases computing power consumption. Moreover, each task runs in an independent cloud virtual machine, which, while ensuring task independence and security, also means that each virtual machine needs to allocate a certain amount of computing power resources, further exacerbating overall computing power consumption.

In this context, the cooperation between Manus and the Tongyi Qianwen team is particularly important. The QwQ-32B model from Tongyi Qianwen, with its excellent reasoning and execution capabilities, can provide strong technical support for Manus; at the same time, the efficiency of QwQ-32B also helps reduce Manus's pressure on computing power consumption, improving response speed and processing efficiency to meet the growing user demand.

Image source: Weibo

This statement shows that the cooperation between Manus and the Tongyi Qianwen team mainly focuses on several key points: based on the Tongyi Qianwen series of open-source models, realizing Manus's functions on domestic models and computing power platforms, and providing services for Chinese users.

"Based on the Tongyi Qianwen series of open-source models, realizing Manus's functions on domestic models and computing power platforms" indicates that the core of this cooperation is a deep integration at the technical level.

In particular, the "domestic computing power platform" may refer to Alibaba Cloud's cloud computing infrastructure. Although Manus's original independent cloud virtual machine architecture ensured task isolation and security, it consumed a lot of computing power.

Alibaba Cloud can provide distributed computing resources and higher computing power efficiency, helping Manus reduce costs while maintaining functional integrity. Additionally, using domestic computing power can meet compliance requirements in the Chinese market (such as data localization).

Considering that each invocation by Manus may consume millions of tokens, achieving equivalent performance on domestic models and computing power while controlling costs will be one of the challenges faced by both teams.

Although more details have not yet been fully disclosed, the cooperation between Manus and Alibaba has already surprised many. This collaboration not only demonstrates the potential for cooperation beyond competition in the AI industry but also provides new ideas and momentum for promoting the industry's development.

After Cooperation with Qianwen, Will Manus Go Open Source?

Last Wednesday, after Manus was released, it quickly went viral on domestic tech media and social networks, becoming a hot topic of discussion. The development team showcased Manus's application cases in resume screening, real estate research, stock analysis, and other tasks, demonstrating its potential in handling complex tasks, very close to the ideal AI Agent.

By the weekend, Manus's influence began to extend overseas, with some foreign media and tech bloggers starting to evaluate and discuss it. The product manager at Hugging Face called it an impressive AI tool, and tech mogul Jack Dorsey on Twitter also expressed his admiration.

The team behind Manus is Monica.im, and the founder, Xiao Hong, is a serial entrepreneur. In an interview, Xiao mentioned the "New Era Andy Beal's Law," where model capabilities overflow, and application companies can seize opportunities.

Regarding the division of labor between model companies and application companies, Xiao believes that vertical fields, specific areas, and "dirty work" may be tasks that original manufacturers are unwilling to undertake, providing application companies with an opportunity to seize this window period and offer more professional solutions. As for the upcoming Manus, its positioning is as a consumer-grade product, a mass-market product, with a pricing strategy leaning towards the consumer market. A comparison is made with the AI agent product Devin, which is priced at $500 per month for the programming market, and ChatGPT Operator, which is priced at $200 per month.

From this perspective, the cooperation between Manus and Alibaba Tongyi Qianwen may essentially be filling the market for "dirty work." The Tongyi Qianwen model can provide strong technical support for Manus, enabling Manus to transform this complex technology into a product friendly to ordinary consumers while meeting the diversity of industry demands and providing precise solutions for specific fields.

Collaboration with the Tongyi Qianwen team may help Manus better achieve this positioning.

Additionally, Alibaba has been enhancing its artificial intelligence capabilities and promoting industry progress through open source. Since 2023, the Tongyi team has open-sourced over 200 models, including the large language model Qianwen Qwen and the visual generation model Wanxiang, covering a full range of sizes from 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B, to 110B, as well as all modalities including large language, multi-modal, mathematics, and code, frequently appearing on authoritative lists both domestically and internationally, becoming one of the most important model series in the global open-source community.

On March 6, the same day Manus was released, the Tongyi Qianwen reasoning model QwQ-32B was also released and open-sourced. On March 10, after Manus was unexpectedly "unboxed," Manus co-founder @peakji (Ji Yichao) responded that open-sourcing is a tradition of the team, "There will be a lot of good things open-sourced soon."

Both parties have a tradition of open-sourcing, and in the future, we may expect Manus to take more actions in the open-source arena.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。