❯ Former DeepSeek-VL lead Tony Zhang releases Linger, a desktop agent that can use local models and signed-in apps
The desktop is the starting pointTony Zhang, who led DeepSeek-VL, founded Simplex AI. The team published Linger’s source code and desktop packages for Apple silicon Macs, Windows and Linux. Users can choose Claude Code, Codex or Pi as the runtime. On Mac and Windows the agent can use the user’s everyday signed-in Chrome browser, and it can also click and type in desktop apps.
Mail and meetings feed a shared memoryAnother Me organizes mail, chats and meetings into a device-local wiki that agents can search. It can monitor the inbox, calendar and Slack, surface decisions that need the user and revise its notification rules from feedback. Linger can also assemble agent teams, delegate long jobs to the background and transcribe meetings live before handing follow-ups to Another Me.
How a 12 GB GPU runs the modelPolyStrata keeps the most frequently activated experts on the GPU and places other weights in system memory or on an SSD. In a team test with a 12 GB RTX 3060 and 64 GB of system RAM, the 125-billion-parameter Qwen3.8-Flash-Next generated 45.5 tokens a second once output began. In a separate long-prompt test with 32 GB of system RAM, the first token took about three minutes. Fitting the model, decoding quickly and answering promptly are separate measures.
▮ SIGNALLinger brings work memory and app control to the desktop, where local responsiveness depends on system RAM and prompt length as well as GPU memory.