Physix Frontier · News Briefing Card (QbitAI · Nov 21, 2026)

DeepSeek V1 author Tony Zhang open-sources startup project Linger

KEY FACTS

  • Tony Zhang, former DeepSeek V1 author and DeepSeek-VL lead, has unveiled his startup project Linger.
  • Linger is a Personal Agent for personal computers, with desktop operation as its priority.
  • Linger's core system, Another Me, can read local email and calendars and build a personal knowledge base.
  • Linger reaches 95% accuracy on the LoCoMo long-term memory benchmark.
  • Linger is open source and supports integration with third-party models as well as local LLM optimization.

KEY DATA

95%LoCoMo benchmark accuracy
45.5 tokens/sRTX 3060 12GB VRAM generation speed
179 tokens/sRTX 5090 32GB VRAM generation speed
125BLocally runnable model parameter count

PHYSIX OBSERVATION

Linger pulls the Agent back from the cloud to local, running hundred-billion-parameter models on open source and consumer-grade hardware, directly challenging the pay-per-token cloud model. For users it is the dual lure of data sovereignty and cost, but the actual experience of running 125B on a 12GB graphics card and the ecosystem integration remain tests. The signal is clear: the battle for the personal Agent entry point has begun shifting from phones to computers.

Source: QbitAI report

DeepSeek V1 author Tony Zhang open-sources startup project Linger | Physix Frontier Briefing