QwenPaw: a self-hosted AI assistant for seven chat apps
The AgentScope team ships version 2.1.0 of its assistant: run it on your own machine or in the cloud, with local models starting at two billion parameters.
Illustration: a plain home office at dusk, where a small computer works away under lamplight.
QwenPaw is an open-source personal AI assistant from the AgentScope team that runs on your own hardware or in the cloud and plugs into seven chat apps.
At a glance
- Licensed under Apache 2.0. The repository requires Python 3.11 or newer, below 3.14.
- Chat services named on the project page: DingTalk, Lark, WeChat, Discord, Telegram, iMessage and QQ.
- Local model line QwenPaw-Flash comes in 2B, 4B and 9B sizes, each with Q4 and Q8 quantizations.
- The v2.1.0 release note is dated August 13, 2026. GitHub displays roughly 34,500 stars and 3,000 forks.
- The project page is the only source available. No independent newsroom has confirmed any of these claims.
QwenPaw is an open-source personal AI assistant from the AgentScope team that runs on your own hardware or in the cloud and plugs into seven chat apps. The code ships under the Apache 2.0 license. It asks for Python 3.11 or newer, and below version 3.14.
Meeting users where they already type
Rather than pushing yet another interface, the project attaches itself to messaging apps people already run. The list covers DingTalk, Lark, WeChat, Discord, Telegram, iMessage and QQ, spanning both the Chinese and the Western market. A web console and a terminal interface round out the options.
Installation paths listed on the project page include pip, an install script, Docker, and a desktop application still marked beta. For teams that would rather not host anything, the page points to one-click deployment on Alibaba Cloud ECS, the AgentScope Platform, and ModelScope Studio.
Local models, not a cloud requirement
The model story is the interesting part. Alongside hosted providers — the page names DashScope, OpenAI, Anthropic, Google Gemini and DeepSeek among others — the project ships its own QwenPaw-Flash line in 2B, 4B and 9B sizes, each offered in Q4 and Q8 quantizations. Two billion parameters is small enough for an ordinary laptop.
Ollama and LM Studio can be wired in as well. How capably the smaller models handle the agent work being advertised is not something the repository answers; no benchmark figures were available for this report.
Security claims come from the vendor
The project describes a layered defense and explicitly claims isolation at the kernel level, implemented through Seatbelt on macOS, Bubblewrap and Landlock on Linux, and AppContainer on Windows. Additional controls are described for tools, file access and installed skills.
All of that comes from the project's own description. No independent audit of the sandbox was available to this newsroom, and until one exists a security claim remains a claim.
What this report cannot establish
Only the GitHub project page was available for this article. That leaves us without the second, independent source we normally require behind every number. Every feature, model size and platform named here is the developers' own account.
The popularity figures deserve the same caution. Roughly 34,500 stars and 3,000 forks are platform metrics from GitHub and say nothing about how many people run the software in production. The v2.1.0 release note carries the date August 13, 2026; whether that is in fact the newest release was not clearly marked on the page.
FAQ
What is QwenPaw?
An open-source personal AI assistant from the AgentScope team, released under the Apache 2.0 license, that can run locally or in the cloud.
Can I run QwenPaw on my own computer?
Yes. The project page lists pip, a shell installer, Docker and a beta desktop app. It requires Python 3.11 or newer, below version 3.14.
Which messaging apps does QwenPaw support?
According to the project page: DingTalk, Lark, WeChat, Discord, Telegram, iMessage and QQ. A web console and a terminal interface are also available.
