Linger runs 125B agent on 12GB GPUs
SimplexAI Labs released Linger, an open-source personal desktop agent that leverages PolyStrata's tiered MoE inference engine to run a 125B parameter model locally on 12GB consumer GPUs. It combines proactive background automation, cross-app integrations, and a self-updating local wiki memory system with zero API fees.
Linger demonstrates that consumer-tier MoE offloading has matured from an inference curiosity into a viable foundation for private, always-on desktop agents. Bypassing cloud API tolls while running high-parameter reasoning locally shifts the personal assistant paradigm from hosted subscriptions to local sovereignty.
- –Tri-tier memory offloading in PolyStrata keeps active router weights in VRAM while streaming dormant experts through system RAM and NVMe, achieving ~45 tokens per second on an RTX 3060.
- –Nightly offline consolidation converts inboxes, calendars, and communications into a structured personal wiki graph, reaching 95% on the LoCoMo benchmark without bloating conversational context windows.
- –Proactive background routines and recursive self-improvement loops allow the agent to monitor tasks and automate workflows locally without leaking telemetry.
- –Zero local token overhead removes the recurring API costs that usually make continuous multi-agent background loops cost-prohibitive.
DISCOVERED
1h ago
2026-11-19
PUBLISHED
2h ago
2026-11-19
RELEVANCE
AUTHOR
0x0SojalSec