The Agent Box, which raised $3 million through crowdfunding, aims to completely solve your computing power anxiety.

Author | Zhang Ziyi

Editor | Yuan Silai

The AI hardware track seems to have stepped into a new world overnight.

Last month, investors were still inquiring about the founders of high - level positions in large hardware companies. This month, they have started to look for "the next Mac mini".

The starting point of the frenzy comes from the open - source intelligent agent framework OpenClaw (referred to as "Lobster" in the circle). From the rush to buy Mac minis with large memory to the co - branded hosts of various hardware and software giants, the fear of missing out (FOMO) is sweeping everyone. The hardware that carries AI has now become an entry point that cannot be missed.

The Agent Box, which was more popular in the geek circle before, has suddenly been pushed to the forefront. Simply put, the Agent Box is a dedicated AI device for individual users, and its only purpose is to run large models and autonomous agents locally.

There are already several companies in the market that have launched Agent Boxes, including Pamir, Violoop, Tiiny, etc. Pamir is valued at over $25 million, and the first product of Tiiny AI, the Tiiny AI Pocket Lab, has raised $2.8 million in crowdfunding on Kickstarter. Industry insiders predict that its final crowdfunding amount may exceed $4 million.

The Tiiny AI Pocket Lab weighs about 300 grams and is about the size of a mobile phone. The early - bird crowdfunding price is $1,399. It supports one - click deployment of large models (up to 120B), does not rely on the cloud, servers, or high - end GPUs, and does not incur additional token consumption fees.

Tiiny AI has undoubtedly caught the wave. After all, users only need to spend the money on a single piece of hardware to use "Lobster" indefinitely.

However, Eco Lee, the vice - president and head of commercialization of Tiiny AI, has repeatedly emphasized in an interview that the Tiiny AI Pocket Lab is not specifically designed for Open Claw. It is an AI infrastructure device designed for individuals.

This sounds fascinating and even a bit hard to believe. People must start to imagine what agent - native means and what it aims to achieve. When the only thing restricting our use of AI is the high token cost, how can we break through this limitation?

Tiiny AI tries to provide an answer.

01 What is an Agent Box?

Before understanding Tiiny AI, it is necessary to clarify a new product concept - what exactly is an Agent Box?

In the past year, in order to run open - source large models locally, people have tried various solutions: some used old and discarded computers, and some rushed to buy top - spec Mac minis.

This hard investment is quite high. If users want to run a large model of over 120B locally, they need to buy a PC with nearly 80GB of video memory, and the cost of the whole machine exceeds 50,000 yuan. Even if they choose Apple's Mac Studio (with the 96GB unified memory version), it will cost more than 20,000 yuan.

"Would you be willing to buy a computer just to run large models? Now, when many AI computers worth tens of thousands of yuan start running local large models, the memory and computing power are over - occupied, and you can't even open a web page. Not to mention playing games or watching videos," Eco said.

In addition to the cost of the device itself, as the token price keeps rising, the high and continuous cost of use has also made "local deployment" a rigid need in the industry.

Therefore, in Tiiny AI's product concept, its product must be a dedicated AI device that can support the local large model and intelligent agent to run in the background 24/7. Its design logic is not to replace the user's personal computer, but to serve as an external independent device for terminal devices such as mobile phones, PCs, tablets, or robots to access and call. The system defaults to saving user data, credentials, and workflows locally. Sensitive operations do not need to be uploaded to the cloud unless a stronger cloud model is explicitly required.

In terms of the software ecosystem, the device will have an Agent Store built - in. Currently, it is compatible with more than 50 open - source large models such as OpenAI OSS, Qwen, and GLM, as well as more than 100 intelligent agent development tools such as OpenClaw and n8n.

To build a rich edge - side ecosystem, Tiiny AI plans to launch a model format conversion tool in July this year. In addition to the SOTA open - source models officially supported by Tiiny, users can also download, convert, and import other open - source models and their own fine - tuned models from open - source communities such as Hugging Face, and upload and share them with other Tiiny users.

"I often give users an analogy. The cloud - based large models are like bottled mineral water. They are good to drink, but ordinary users have a large number of high - frequency, repetitive AI needs that are close to personal habits and do not require top - level intelligence. It's too wasteful to use mineral water to wash hands and take a bath. Tiiny is like an 'AI faucet' for users. You can use it at will, and the marginal cost of tokens is 0," Eco told Yingke.

In Eco's view, cloud - based large models focus on handling high - intelligence, high - precision, and critical tasks, while local large models focus on daily high - frequency, personalized, and continuous interaction scenarios with long - term user memory. This "edge - cloud collaboration" model is the core value of Tiiny AI and also the Agent Box.

Amidst the expectations, applause, and a stream of investment invitations, Tiiny AI inevitably has to face doubts. The first question it must answer is how to achieve the cutting - edge product concept and a 120B parameter model at a not - so - expensive price?

02 Toy or Tool?

On the overseas reddit forum, the reviews of the Tiiny AI Pocket Lab are polarized. Some people say it will only be a toy, and some have reverse - engineered Tiiny AI's product through the promotional photos and believe that the functions it claims are difficult to truly achieve.

The point they question is that Tiiny AI has not announced the brand of the SoC (system - on - a - chip) it uses, and it does not use high - end GPUs, yet it can run a 120B large model locally.

This is quite incredible.

"We are an AI Infra company. The core is to optimize the limited hardware's computing power and resources through systematic underlying optimization, and focus all of them on LLM inference and agent operation. This is fundamentally different from the ideas of other hardware companies," Eco said.

The chip used in the Tiiny AI Pocket Lab is an SoC plus a dNPU, and through Tiiny AI's core technology, PowerInfer, it can achieve local model inference capabilities comparable to high - end GPU chips such as Nvidia and AMD.

PowerInfer is a heterogeneous computing power inference acceleration technology for the edge side. The Tiiny AI team found through a large amount of data calculation and corpus training that in the process of large - model inference, the parameter activation modes are divided into two types: "hot - activated parameters" (core parameters that are called every time when interacting with the model, accounting for about 20%) and "cold - activated parameters" (activated only when users ask questions in specific fields such as medicine and law, accounting for about 80%). This characteristic of hot and cold activation is just suitable for optimization and allocation under the edge - side heterogeneous computing power architecture. The team has open - sourced an example of PowerInfer: using a single NVIDIA RTX 4090 GPU to run a large model with 175B parameters, the speed can reach 11 times that of the traditional solution.

These all belong to the technical accumulation at the AI Infra level. From the chip layer to the agent scheduling layer, and then to the model training layer, all require in - depth know - how support.

In specific implementation scenarios, the Tiinny AI team found from the message area of Kickstarter that its users are mainly ordinary users who use open - source applications such as OpenClaw, professionals with a rigid need for data privacy, and AI geeks. Even in an offline environment without an internet connection, the device can still run multi - step inference, agent workflows, content generation, and secure computing for sensitive data. The "fool - proof" out - of - the - box usability, 24/7 agents with 0 token fees, and complete autonomous control are the core reasons for these people to choose Tiiny.

Moreover, the device has introduced a long - term memory function into the system. Users' interaction preferences, historical files, and conversation records can all be directly stored in the local hardware in an encrypted form.

"Privacy is an added bonus for the Agent Box, but the core lies in the deployment of the local model. It can actively do things in combination with your long - term memory. This is the most important thing," Eco told Yingke.

Tiiny AI's products are expected to be delivered in August 2026. It should be noted that Tiiny is a team with a background in AI agents, and its final product is still hardware. They have cooperation suppliers, but they also need to deal with unexpected situations in mass production.

There are too many challenges in hardware production. What tests the team is not financing, but truly fulfilling the attractive promise: to achieve local computing power freedom, be free from the constraints of token prices, and ensure complete privacy on a 300 - gram box.