A new open-source personal AI agent called Linger says it can run a 125-billion-parameter model on a PC with a 12GB graphics card, keeping the work on the user’s own machine. If its reported speed holds up, that is a striking attempt to make large-model inference less dependent on cloud accounts and token bills.
Watch Desk analysis
What happened
Linger is a PC-focused agent from Simplex AI, founded by former DeepSeek contributor Tony Zhang, according to 36Kr. It is designed to work with logged-in apps and browser pages, including email and calendars, and can pass control back to the user when a task gets stuck. The project is open source and supports third-party models.
The report says Linger’s PolyStrata inference engine distributes a model’s components across GPU memory, system RAM and storage. Its official figures put a 125B-parameter model at 45.5 tokens per second on a PC with an RTX 3060, 12GB of VRAM and 64GB of RAM. That is a claim about a particular setup, not a promise that every 12GB card will produce the same result.
Linger also describes a local knowledge base built from sources such as inboxes, calendars and meeting notes, plus an agent that can learn reminder preferences. The report says its system scored 95% on the LoCoMo conversational-memory benchmark. Those capabilities and results are presented in the 36Kr account; the benchmark score is attributed to the project.
Why it matters
The headline is not simply that another assistant can read your calendar. Linger is trying to make local inference practical by moving model components between different kinds of storage, rather than requiring the entire model to fit in scarce graphics-card memory. If the reported performance is reproducible, more people could experiment with capable models on hardware they already own.
That comes with a trade-off: the example requires 64GB of system RAM, and reported speeds and benchmark scores are not the same as an independent comparison. Still, the idea of an agent working with locally held data, without sending every task to a cloud service, is a useful direction to watch. Your computer may need a bigger memory budget, but at least it would not be asking for a monthly token allowance.
Our read
This is a substantive local-AI project, not just a fresh coat of chatbot paint. The combination of desktop control, local data and an inference engine designed around ordinary consumer hardware is ambitious. We would treat the performance figures as a starting point for testing, not a verdict: reproducible comparisons, hardware requirements and clear permission controls will matter as much as the demo.
What to watch
- Whether independent users reproduce the reported model speed and memory requirements.
- Which models and hardware combinations Linger supports in practice.
- How it asks permission before reading personal data or taking actions in logged-in apps.
Discussion spark: Would you let a local AI agent work across your email and calendar if it kept your data on your own machine, or should that level of access remain off limits?
Sources and evidence
- DeepSeek Team Member Zhang Bo’s New Startup Exposed: Launching Open-Source Personal AI Agent – 36Kr (20 November 2026, 01:22 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.