DeepSeek V1 Author Reveals Open-Source Local Agent
Tony Zhang’s Linger targets a different agent battleground: a PC-native, open system that keeps memory, data and execution on the user’s machine.
What launched
Tony Zhang, identified as a former DeepSeek V1 author and DeepSeek-VL project lead, has revealed Linger, an open-source personal agent developed through his startup Simplex AI. The project is designed to run on a user’s own computer rather than primarily through a mobile app or cloud service.
Linger can work with existing desktop environments, including browsers, email and calendars, while allowing users to intervene when an agent gets stuck. It is also positioned as a model-neutral layer, able to connect to different commercial and local models. The project says its local inference stack can run a 125-billion-parameter Qwen3.8-Flash-Next model on an RTX 3060 with 12GB of VRAM, reaching 45.5 tokens per second; that claim still needs independent testing.
Why the architecture matters
The product’s most ambitious component is “Another Me,” a memory and orchestration system that continuously organizes inboxes, calendars, chats and meetings into a local knowledge base. Linger says it can learn notification preferences from user feedback and coordinate multiple specialized agents behind a single session. It reports 95% accuracy on the LoCoMo long-term-memory benchmark, but the comparison setup and reproducibility details are not yet clear.
This approach differs from cloud-first assistants that treat the user’s context as a service-side resource. Linger is betting that privacy, predictable local costs and direct control over permissions will matter enough to outweigh the convenience of hosted agents.
Why it matters
The announcement is not a frontier-model release, and the project’s real maturity remains unproven without a public repository, reproducible benchmarks and sustained reliability data. But it is strategically notable: a researcher associated with DeepSeek’s model era is now pushing the competition toward local agent infrastructure, where memory, permissions and desktop control may become more important than another leaderboard score.
Uncle Cat take
Linger’s 12GB-VRAM claim is the story’s sharpest edge, but until outsiders reproduce 45.5 tokens per second, its local-agent thesis remains an attractive specification rather than a demonstrated advantage.