Reportedly: Simplex AI founder Tony Zhang open-sources Linger, a personal agent that can run locally
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Simplex AI founder Tony Zhang announced that he had open-sourced Linger, a personal AI agent that can run on a user's computer with a local model, Zhidongxi reported. The report said Linger passed 2k GitHub stars and cited project-reported accuracy of 95% on the LoCoMo memory benchmark.
China context
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- The report points to a GitHub repository but does not verify a working installation outside China. Can users there install Linger and run it with a local model without relying on a region-restricted service?
Translated from Chinese. Quotes and facts link to the original sources.
Zhidongxi reported that Linger's Another Me assistant monitors inboxes, calendars and Slack for tasks, sends reminders and uses feedback to adjust when it reaches out. The outlet also reported that Linger can coordinate multiple agents, work with local files, browsers and desktop apps, and connect to Claude Code and Codex.
Zhidongxi also reported that Simplex AI open-sourced PolyStrata, a local inference engine. Zhang said it could run the 125B-parameter Qwen3.8-Flash-Next model with 64 GB of system RAM and 12 GB of GPU memory, and reported 45.4 tokens per second on an RTX 3060 12 GB. He separately reported 80 tokens per second with frequently used model experts in GPU memory and less frequently used experts in system RAM. These figures do not establish end-to-end performance for Linger; a reproducible test would need to specify the model, quantization, context length and hardware for each run.
Linger users who run a local model would not pay cloud token fees for inference performed on their own hardware, if the reported setup works as described. That does not establish lower total costs or fully offline operation: an independent test should check which workflows still make network calls when Slack, email or external coding tools are enabled.
Teams evaluating local agents can compare hardware and maintenance costs with their current API bills for the same workload. GitHub stars indicate early interest, not sustained use or savings.
The next checks are whether independent testers reproduce the LoCoMo score and inference speeds, and whether they can complete a representative task using only a local model with cloud model services disabled.