Chinese researchers’ AI prowess is again a hot topic after a startup called Monica.im last week revealed “Manus”, a service it bills as a “general agent” that might improve on tools offered by Western companies.
Manus is being compared to OpenAI’s Deep Research, which scours online services to source info that’s compiled into documents that OpenAI claims “create a comprehensive report at the level of a research analyst” within half an hour. Another point of reference is tools like Anthropic’s Computer Use API and OpenAI’s Operator Agents, both of which can use a web browser to perform basic tasks like filling in forms and using e-commerce sites.
Manus looks like it does all that and more – maybe faster too according to its own benchmarks. A launch video depicts it doing three chores at super-speed:
- Recommending the best candidate for a job after ingesting, opening and reading job applications, then ranking candidates in a prose document before – upon being prompted to do so – reformatting its recommendations as a spreadsheet;
- Preparing a report on available properties created after a user offers info on budget, requirements, and desired location. The report includes available listings, plus info on neighbourhood amenity;
- Conduct correlation analysis of different stocks, write a prose document with conclusions and create an interactive website that lets users explore data scraped from the web.
Manus offers the familiar chatbot user interface of an empty text field into which to type a prompt. Early testers have described the experience of using Manus as akin to sitting with someone who is sat at a keyboard and turns vague instructions into precise output at extraordinary speed.
The service runs in “Manus’s computer” – which appears to be a cloudy Ubuntu workstation. The launch vid states that the service “operates as a multi-agent system powered by several distinct models”, some of which will be open sourced later this year.
Demos of the service shows that workstation writing its own commands, visiting websites galore, and then delivering a document and the complete code used to produce it.
The Register is not impressed by some results. A demo of a Mario-style platformer game created by Manus is crude and crashed. An itinerary for a two-month trip to “Australia, then New Zealand, Argentina (and other parts of South America), and Antarctica” cites just 17 sources for its output, and suggests flying in “luxury” will cost only double the price of budget fares when Business Class nearly always costs at least three times economy fares.
The holiday plan also fails to produce a promised full downloadable , and makes odd suggestions for a month spent in Australia: Who comes for a month without visiting Sydney, but goes to Tasmania in winter and spends a week in parts of the Outback that need three days at most?
We’re already seeing reports of slow performance and unsatisfying output.
But we’ve found plenty of testers who’ve had happier experiences and report the tool has opened 50 browser windows at a time to source data, then analyzed it in a flash. Developers have marveled at its coding abilities.
A lot of commentary we’ve seen latches onto the term “general agent” used by Monica.im and jumps from there to describe Manus as a step towards artificial general intelligence – software that can perform tasks with human or superhuman skill.
DeepSeek 2.0? DeepSigh, more like it
We’re also seeing plenty of commentary that suggests Manus represents a leap in performance like that achieved by Chinese startup DeepSeek, which created a chatbot that produced fine results and was initially thought to require much less compute power than rival services.
That assumption led to something of a panic as it was assumed Chinese AI companies were beating all rivals despite export bans designed to prevent that from happening. Investors also paled as they contemplated hyperscalers’ planned mega-spend on AI infrastructure and whether it could be unnecessary and therefore hard to recoup.
Both reasons for panic were eventually debunked, as DeepSeek was found to have lousy security that saw it banned by several governments, and to have overstated its claims about requiring only modest hardware. It’s also biased: Prompts that would produce a response that show the Communist Party of China in a poor light produce errors.
Hyperscalers told investors not to fear, because they’re building infrastructure for inferencing and expect that to become a part of almost every application in coming months and years (and are hiking prices for apps that rely on it).
The Register has applied for an account on the invitation-only service and hopes to be accepted so we can offer a hands-on evaluation in coming days. ®
