Manus Is Said to Train Its Own Model After $500M Round

Manus, the AI agent start-up, is training its own model, 36Kr reported on 24 October 2026. “We learned that Manus is also training its own model,” 36Kr’s English edition wrote, in an interview piece on another personal-AI start-up, Today AI. The report gave no model size, cost or timeline, and it did not name its source. We found no comment from Manus.

The report comes weeks after Manus’s parent company, Butterfly Effect, announced a round of over $500 million on 8 October, according to 36Kr. Today AI, the start-up 36Kr was writing about, announced over $200 million on 21 October, and its team says its vision includes training models of its own.

Business Pill · THE MODEL AND THE HARNESS

A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and says nothing about Manus or Today AI.

What 36Kr Reported

The 36Kr piece is an interview with the founding team of Today AI, a personal assistant start-up led by Qi Junyuan, who founded Teambition. 36Kr says Today AI has launched international and domestic versions since 9 September.

Its first round of over $200 million, announced on 21 October, drew IDG Capital, Tencent, 5Y Capital, Ant Group, HSG, Transsion Holdings, Tongge Venture Capital and Hanyuan Asset of Shanghai Jiao Tong University, 36Kr lists. The team told 36Kr that “their vision is not only to build AI applications, but also to train models”.

36Kr then widened the frame: “This is a common trend among recent star AI application companies.” It noted that Manus launched its personal agent product, Cue, on 28 September and then announced its round, before adding the line on Manus training its own model.

What Was Posted on X

The X account @choblin29 flagged 36Kr’s report on 24 October, writing that Manus built its agent business around other labs’ models rather than training its own foundation model, and that Today AI is training a dedicated ‘interaction model’. We checked the report against 36Kr’s English edition, which describes Today AI’s plan as a stated vision to train models.

Why an App Company Would Train a Model

36Kr gives its own reading of the move. It writes that “application companies move upstream to train models to reduce unit intelligence cost, improve effect and gross profit”, while model companies move downstream to build applications.

In plain terms, a company that rents another lab’s model pays for every task its agent runs. Owning a model, even a small one, is one way to cut that cost per task. Today AI says its own engineering work includes “using as small a model as possible”.

The team was also open about demand. “It is still too early to draw the conclusion that we have reached a very mature PMF stage and users are already willing to pay for it,” it told 36Kr, adding that it is still verifying this. PMF stands for product-market fit, the point at which enough customers want a product and pay for it.

Manus’s Road to a $500M Round

An earlier 36Kr article sets out how Manus got here. Meta announced the acquisition of Manus for more than $2 billion in December 2025. In April 2026 a regulator required the deal to be revoked, the original investors repurchased the shares from Meta, and on 1 September Manus announced that it had resumed independent operation.

Boyu Capital and IDG Capital led the 8 October round, with Tencent, HSG and ZhenFund adding to their holdings, according to 36Kr. It called the round “the largest single financing in the domestic AI application sector to date”. The target valuation was about $4 billion, it wrote, but Manus did not disclose the final valuation.

36Kr also cites a June report by The Information: Manus’s annualised revenue run rate was about $100 million when Meta’s acquisition was announced in December 2025, and $400 million to $500 million by June 2026. Those are reported figures, not company disclosures.

On the model side, 36Kr wrote that Manus is reorganising its China team, “promoting cooperation with domestic large model vendors” and recruiting in Beijing.

The Rounds Side by Side

Three personal-agent start-ups announced large rounds between 28 September and 21 October. The table lists each round as reported and what each has said, or been reported to say, about its own model.

Connecting the Coverage

Our Manus 2.0 piece of 29 September covered the launch of Cue and the Cascade agent framework, and found the company’s blog disclosed no model specification. Our valuation piece of 19 September covered The Wall Street Journal’s report that Manus was seeking $500 million at about $4 billion. Our June piece covered China’s order to unwind the Meta deal.

Other agent and app companies have been edging toward the model layer. Our Cursor piece of 9 August covered a leaked checkpoint, unconfirmed by Cursor, that pointed to a pre-trained coding model of its own. Our Mubit piece of 23 October covered a service that trains small models on a company’s own agent calls, at $5 per 1,000 runs.

The opposite view came from Microsoft. Our Nadella piece of 21 October quoted his shareholder letter: Microsoft separates the harness from the model, so that “models become interchangeable.”

What’s new: the first report we have read that Manus is training a model of its own, alongside a second personal-agent start-up, Today AI, whose team says its vision includes training models.

The Structural Read

As we read it, a personal agent that rents a frontier model pays the model maker on every task, so its margin depends on someone else’s price list. Training even a small model of its own is one way to keep more of each task, which is how 36Kr frames it.

As we read it, this also cuts against the swappable-model view in our Nadella coverage. Microsoft says it separates the harness from the model so that models become interchangeable; Manus and Today AI, on 36Kr’s account, want a model of their own under the harness.

As we read it, the report is thin on Manus itself: one sentence, no source named and no detail on size, cost or timing. Today AI is clearer on intent, because its team told 36Kr that its vision from day one includes training models, and it also says product-market fit is still unproven.

Three Implications

FOR MODEL LABS 36Kr writes that model companies move downstream to build applications to find revenue outlets, and that “the boundary between AI applications and models will naturally become increasingly blurred”.

FOR INVESTORS IN AGENT APPS Manus’s round of over $500 million comes with a 36Kr report of in-house model training, and Today AI’s of over $200 million with a stated vision to train models; neither company has given figures for what that training costs.

FOR ANYONE BUILDING ON RENTED MODELS Today AI told 36Kr its engineering includes “using as small a model as possible”, and Mubit, in our earlier coverage, sells small models trained on a company’s own agent calls.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework puts it this way: “Everything downstream is a consequence of value migrating from the thing you talk to toward the thing you delegate to.”

As we read it, Manus and Today AI already sell the thing you delegate to, a personal agent. 36Kr’s report suggests they now also want to own the model underneath it. The framework was published in July 2026 and does not discuss Manus or Today AI.

What Is Not Established

36Kr’s Manus report is one sentence. It gives no model size, compute, cost, team or release date, and it does not say whether the model would replace the outside models Manus uses. We found no confirmation from Manus or Butterfly Effect.

Manus’s final valuation in the 8 October round was not disclosed. The run-rate figures come from The Information as cited by 36Kr. Today AI did not give details of its own model in the piece we read.

The Bottom Line

Manus, which announced over $500 million this month, is training its own model, 36Kr reports, and Today AI, which announced over $200 million, says its vision includes training models of its own. 36Kr reads it as application companies moving upstream to cut the cost of each task. Manus has not confirmed it.