Violoop Secures Series A Funding to Develop 'Self-Evolving Personal AI' Hardware

In 2026, competition in the AI office sector intensifies: Tencent's WorkBuddy reaches 20.97 million visits within three months of launch, while ByteDance integrates Feishu and Doubao services. As major tech firms vie for the desktop AI workspace entry point, startup Violoop secures funding in the hundreds of millions, entering the AI Agent market through hardware. Violoop develops a handheld external hardware device, roughly the size of a gaming console, capable of connecting to a computer to access screen data and execute tasks. Its core innovation lies in the “proactive AI” concept, leveraging long-term memory, contextual understanding, and task pre-preparation to complete work before the user speaks. To address security concerns, Violoop employs a layered architecture featuring on-device processing for sensitive tasks, an independent security chip for authorization management, and physical buttons to confirm irreversible operations.

Article author and source: IT Juzi

In the summer of 2026, "AI helping workers get things done" is moving from concept to real-world applications.

According to public data from Analysys, in June 2026, Tencent WorkBuddy, after only three months of launch, ranked first among domestic desktop-based AI-native office agent platforms with 20.97 million visits.

ByteDance is also rapidly integrating its office AI business. On July 30, the Feishu product team merged with the Doubao product team; on August 25, the productivity-focused Agent product "Doubao Work" was officially launched and deeply integrated with Feishu. Users can state their goals in natural language, and the system will break down tasks, invoke tools, and advance workflows accordingly.

Big tech companies are competing for the same spot: the AI workspace entry on the desktop.

When models, traffic, and application ecosystems are all controlled by large corporations, where are the opportunities for startups?

IT桔or has learned that Shenzhen-based AI hardware company Violoop has recently completed its billion-yuan seed and Pre-A financing rounds, led by Lenovo Capital, CICC Porsche Ventures, Blue驰 Ventures, YuanSheng Capital, Qifu Capital, and Zhenke Holding, with Xiangyang Capital serving as the long-term exclusive financial advisor.

However, Violoop did not directly compete with big tech companies; instead, it entered the AI hardware market by developing an external hardware device the size of a handheld console that can connect to a computer.

It features its own dedicated chip and edge computing power, can connect to a computer via HDMI and USB, receives visual input from the screen, and executes actions through keyboard and mouse signals. In other words, Violoop aims to create a physical-level “personal AI agent.”

In light of the good news about funding, IT桔子 spoke with He Jialin, the founder of Violoop (a millennial), about the product’s strategic direction. The choices regarding its hardware form factor and software logic all ultimately point to three key questions: Where is the upper limit of AI capability? How can we ensure the minimum threshold of AI trust? When software agents are already powerful enough, why still need a hardware device?

Photo source: Provided by the interviewees He Jialin, founder of Violoop, and Zhu Xianzhen, co-founder and CTO

The mainstream AI agent assistants today still operate on a "you ask, I answer" model.

User requests are initiated by the user, understood and executed by AI, after which it pauses and waits for the next instruction. Even as some products begin to incorporate calendar reminders, scheduled tasks, and automated triggers, they still fundamentally rely on users setting rules in advance.

Violoop aims to create another form of "active AI."

Proactive AI continuously understands what the user is doing and completes necessary preparations before the user asks, based on the current context, long-term memory, and past working patterns.

During early discussions and product testing with multiple users, the team identified a common issue: what often hinders AI from entering real-world production work is not just the model’s capability, but the prompts themselves.

To enable an AI to complete a complex task, users typically need to pause their work, reorganize the context, locate relevant materials, define their goal, break down the steps, and condense the scattered context—from screens, files, history, and personal experience—into a clear, structured prompt that the machine can understand.

For a small number of skilled users, this is a technique; but for most ordinary users, it is itself an additional task. When the cost of delegating a task approaches the cost of completing it personally, agents struggle to truly enter high-frequency, continuous production environments.

Therefore, Violoop aims to eliminate this "pre-task"—the requirement of providing output prompts as a necessary prerequisite for the AI to understand the task and begin working.

As a result, a new interaction paradigm has emerged: the system understands what is happening in the moment by leveraging continuous context, long-term memory, and real-world workflows, and completes information organization, task decomposition, and reversible preparatory actions before the user actively speaks up.

Of course, the prompt won't disappear entirely, but it will serve as a necessary supplement and calibration tool rather than a mandatory starting point for every task.

Violoop calls this capability Artificial Intuition.

In the beta version demonstrated live by He Jialin, this experience typically consists of several consecutive scenarios:

When a user is discussing something with a colleague on WeChat, the system has already prepared the next reply based on the current conversation and relevant context, allowing the user to accept it with a shortcut key; when a colleague requests a candidate’s resume during the chat, the system understands the need, locates the corresponding file, and pauses before sending, waiting for the user’s confirmation; when the user opens the resume to continue reading, the system, in conjunction with the current project, asks whether a job matching analysis is needed and pre-organizes the required information.

In other words, the user doesn’t need to explicitly open an AI app, re-enter context, copy chat history, or write prompts—the system has already advanced the task to the point where it’s waiting solely for the user’s decision.

Behind this lies not just a smarter model, but more importantly, the construction of a system that includes cross-app awareness, long-term context, personal memory, and the ability to judge user intent and permission boundaries.

How exactly do you do it? He Jialin breaks down Violoop’s learning path into three stages.

In the first phase, the AI develops long-term memories centered around the individual.

The system gradually learns about the people, projects, files, and historical work associated with the user from their authorized workflow context. It determines which information is worth preserving long-term and which is merely short-term noise, while continuously merging duplicates and correcting outdated information.

The second stage is to teach the AI how a person works.

The system must not only remember what has occurred, but also gradually understand where users seek information, how they handle similar tasks, how they make decisions, and which steps can proceed versus which require confirmation.

These repeatedly used and refined approaches gradually solidify into skills, workflows, and execution paths, entering the Agent Harness—the Agent’s execution framework.

The third stage is to transform repeatedly validated experience into a more stable personal model capability.

Violoop has established the technical pipeline of "capturing work trajectories—forming long-term memory—enabling personalized training" and is currently undergoing engineering implementation. The team is simultaneously training and testing two types of proprietary models: one focused on task understanding and task decomposition, and the other on concrete execution.

He Jialin compared this change to "muscle memory."

An experienced person doesn’t need to reason from scratch every time they face a familiar task—years of experience have been condensed into rapid judgments. Violoop hopes that agents can similarly gradually accumulate real-world outcomes, feedback, and corrections over time, rather than starting from zero with every task.

“What large models lack now isn’t just intelligence,” He Jialin said. “What’s even rarer is how to preserve the experience from every real-world task and ultimately form a personalized workflow for each user.”

However, for proactive responses to be effective, there is another very practical technical requirement—timeliness.

If the system provides suggestions only after the user has already left the current page, even accurate suggestions lose their value. To address this, Violoop has set response targets for key proactive interactions to under one second, increasing edge AI processing power from 6 TOPS in the previous generation to 26 TOPS (a unit measuring an AI chip’s computational capacity per second). For comparison, the maximum processing power of a mainstream AI PC from a major manufacturer is 40 TOPS—similar to the performance of a mini desktop computer.

This significant boost in edge-side computing power will enable Violoop to achieve lower-latency perception while also handling memory processing, local model inference, and certain personalized tasks.

In April this year, an AI programming agent, while handling a task in the testing environment, obtained overly privileged API credentials and deleted a startup's production database and its volume backup within seconds.

Fortunately, the relevant data was later restored, but the incident revealed a deeper concern: how can we ensure that an Agent does not perform unauthorized actions or accidentally delete critical data when granted execution rights? How can we stop it if it misunderstands or deviates during execution?

The stronger the active AI capabilities, the more security becomes an unavoidable and necessary issue to address.

Violoop's response to this issue is divided into two levels.

Layer one: Prioritize tasks closest to the user—those most sensitive to latency and privacy—to be completed locally.

Violoop handles on-device tasks such as screen awareness, memory construction, partial text completion, and low-latency model operations; more complex tasks requiring stronger reasoning capabilities are dynamically routed to cloud models as needed.

Although hardware itself does not automatically ensure privacy, standalone devices can separate perception, memory, model inference, and cloud-based calls for complex tasks into clearer, more controllable data boundaries.

Moreover, since users have varying preferences for large models, they should not be forced into a specific model package. Users who already possess API keys for models such as Zhipu, DeepSeek, or others can integrate these models themselves; the primary value of the official model packages lies in centrally providing and managing multiple models, allowing users to make choices based on performance, speed, and cost.

Layer two: Separate critical authorizations from the model and general software.

Violoop is equipped with an independent security chip that manages authorization on the execution side. When the model encounters operations defined as irreversible—such as sending, deleting, submitting, or paying—it must pause and wait for the user to confirm via a physical button before proceeding.

This architecture does not guarantee that AI will never make mistakes. Instead, it ensures that any irreversible action included in the authorized workflow can never proceed without final human confirmation.

He Jialin used the analogy of “signing a contract”: the terms of the contract can be revised many times, and documents can be prepared in advance by others, but the moment the seal is stamped, the decision becomes officially effective.

"An agent can help users complete the first 80%, or even 99%, of the preparatory work, but the final step should still be confirmed by a person," said He Jialin. "We want to keep the authority to sign and stamp always in the hands of the user."

If a system can only provide suggestions but cannot interact with real software to perform actions, its value is limited; however, if it can freely send messages, delete files, or complete payments without restraint, users will find it difficult to trust it over the long term.

In Violoop’s product logic, capability boundaries and trust boundaries must grow in sync; AI can become increasingly proactive, but decision-making authority must not be transferred as a result.

At the end of the interview, I asked whether pure software can truly not achieve things like long-term memory, computer use, local models, and security verification.

He Jialin's response was that, individually, these capabilities are not exclusive to hardware; however, in Violoop’s chosen technological approach, hardware integrates several previously disparate capabilities into a single, cohesive product.

The primary purpose of hardware is to remain continuously available.

Active AI needs to know what is happening right now and also connect the current state with past memories.

Violoop captures the same screen information as the user via HDMI and continuously processes authorized context through a dedicated device. It does not depend on whether a specific application is open, nor is it fully limited by the lifecycle of a single software session.

Hardware enables constant presence, and constant presence enables complete context. For Violoop, this forms the foundational data for long-term memory and active judgment.

The second reason is limited software coverage and restricted compatibility.

Pure software agents typically rely on operating system permissions, software interfaces, or API adaptations. However, in the real-world software landscape, many applications lack standardized interfaces for AI, and numerous legacy systems cannot be individually modified for every type of agent.

Violoop uses interaction methods that humans have been using for decades: obtaining information from the screen and performing actions through standard keyboard and mouse interfaces.

This means its scope of capabilities does not need to be entirely dependent on whether the software has a dedicated AI API; theoretically, any software that can be operated via screen and keyboard/mouse input has the potential to be included in its working range.

The third reason is to establish a physical trust boundary independent of the model.

If perception, reasoning, execution, and authorization are all contained within the same software system, a failure in the model, plugin, or permission configuration could cause risk to propagate along the same chain.

Violoop attempts to separate these capabilities: the AI model handles understanding and preparation, the execution system handles operations, and independent security chips and physical buttons handle final authorization for irreversible actions.

Software enhances AI capabilities, while hardware determines data flow, execution boundaries, and where the system must stop.

These three points together constitute what Violoop understands as "AI-native hardware."

Their original goal in developing this hardware device was to enable AI to continuously understand and use computers—every aspect, from selecting computing chips and input/output interfaces to local models and security architecture, serves this core requirement.

Violoop’s core team primarily comes from software and model engineering backgrounds. From 2023 to 2024, the team provided model deployment, post-training, and optimization services on computing clusters for European Fortune 500 companies.

Now, they are compressing AI systems that previously required the computing resources and customized delivery of large enterprises into desktop devices accessible to ordinary consumers.

On the product side, the first batch of Violoop beta hardware has received ongoing feedback from dozens of real users, and broader product testing has encompassed over 200 users from multiple countries and regions worldwide.

According to the team’s current plan, Violoop will officially launch on overseas channels in mid-September, with an initial small-scale release in China, followed by gradual expansion of market coverage after October.

While major companies compete for AI entry points in software, Violoop chose a path that appears heavier and more difficult.

It needs to demonstrate whether the hardware can deliver more complete context, lower interaction latency, and more trustworthy execution boundaries when the AI Agent begins operation.

If these three things can be validated at scale, the interaction paradigm for personal AI agents may be redefined.