Simplex AI has made the source code for Linger public—a personal agent designed to run on a user's own computer using local language models. The project was featured in Chinese publications on November 20th. The project founder, Tony Zhang, previously led the DeepSeek-VL project.
Linger is a product of Simplex AI, not a release from DeepSeek. The system's central agent, named Another Me, is designed to structure information from emails, calendars, chats, and meetings into a unified user knowledge base. It can also identify tasks requiring follow-up and delegate them to other agents. The project description also mentions browser and desktop operation capabilities with the option for human intervention when necessary. It should be noted that these functions are described by the developers and have not undergone independent product testing.
The local inference component, called PolyStrata, distributes the loads of 'mixture-of-experts' type models across the Graphics Processing Unit (GPU) memory, System Random Access Memory (RAM), and SSD storage, instead of requiring all experts to reside on the video card. More frequently used experts remain closer to the GPU, while less demanded parts can be served through other memory levels.
Simplex AI reports that the Qwen3.8-Flash-Next model with 125 billion parameters generates 45.5 tokens per second on an RTX 3060 with 12GB of VRAM and 64GB of RAM, and 179 tokens per second on an RTX 5090 with 32GB. These figures are manufacturer-reported decoding speed measurements and do not represent full agent task time characteristics or guarantees of equivalent quality to cloud models. Additional hardware planning tables include budget memory configuration simulations, which limits comparisons between different machines.
The public release of the code does not imply unlimited licensing. The main license modifies Apache 2.0 by adding additional conditions: running Linger as a hosted or managed service for third parties requires written permission from the manufacturer, and removing or changing the frontend branding is only possible with a commercial license. The terms allow for internal commercial use and use as a backend for the organization's own applications.
At the time of checking, the main website indicated that downloads for macOS, Windows, and Linux as ready-made packages are expected in the future, despite the fact that the source code and setup documentation were publicly available. Code availability should not be confused with the release of a finished installer. Furthermore, local inference does not guarantee fully autonomous operation, as connected applications, voice commands, remote models, and background tasks using external subscriptions may transmit data to external providers. The security documentation describes server assertions and checks, not a sandbox isolating agents from owner files. Pandaily did not install the software nor reproduce the claimed performance.



