Running a Personal AI Agent on One RTX 4090: Open-Source Linger Plus PolyStrata, 5x Faster Than llama.cpp
lxfater · x · 2026-11-19
A developer shares how he built a personal AI assistant on a single RTX 4090 using the open-source Linger agent and its PolyStrata inference engine — his own model choices, unlimited tasks, zero token cost.
PolyStrata (inference engine)
- Layers MoE experts across storage tiers: hot experts in VRAM, cold ones in RAM/SSD — runs on laptops with just 8GB VRAM
- Supports 12 models (Qwen, DeepSeek, GLM, Gemma, etc.) with hot-swapping; up to 5x faster than llama.cpp on the same hardware
- OpenAI/Anthropic-compatible APIs, so Claude Code and Codex plug in directly
Linger (personal agent)
- Connects email, chat, and calendar for cross-app context; email drafts require human confirmation before sending
- Proactively acts: books time and writes briefings when a client asks about budgets
- Spawns an agent team for bigger jobs and hands results back for review
Models are freely mixable — a frontier model as the orchestrator via existing subscriptions, local models for the rest. The author runs Linger on Mac against a locally hosted Gemma on the 4090.
More from coding & agent
- Replit MCP Server lets AI assistants manage workspaces, files and deployments — modelcontextprotocol · 2026-11-19
- Open-source t3boat runs T3 Code agents on cloud machines that sleep when idle — gethackteam · 2026-11-19
- CS Professor Lemire Slams AI Slop in GitHub Issues: 'Answer Me, Don't Throw Slop at Me' — lemire · 2026-11-19
- Univer hits #1 on GitHub Trending: open-source AI office SDK with 29.9k stars — alifcoder · 2026-11-19
- User builds tower defense game in 20 minutes with Gemini 4 Neon, using only 2% of quota — DeryaTR_ · 2026-11-19
- ProtoInputMultiPlex: an MCP server giving AI agents their own virtual mouse and keyboard on Windows — Kenzillla · 2026-11-19