simplexai-labs/linger

Linger, developed by SimplexAI, offers two installation methods: a desktop application and Docker deployment. Its PolyStrata engine allocates VRAM based on expert usage frequency. The "Model Autopilot" feature detects hardware, recommends, and installs models with a single command. The "Another Me" feature proactively follows up by reading inboxes, calendars, and Slack, rewriting its own decision rules nightly based on your reactions. Agents can also operate applications on the computer and logged-in Chrome instances, form teams, run long tasks in the background, and switch to using Claude Code, Codex, or Pi as the runtime.

Linger, developed by SimplexAI, offers two installation methods: a desktop application and Docker deployment. Its PolyStrata engine allocates VRAM based on expert usage frequency. The "Model Autopilot" feature detects hardware, recommends, and installs models with a single command. The "Another Me" feature proactively follows up by reading inboxes, calendars, and Slack, rewriting its own decision rules nightly based on your reactions. Agents can also operate applications on the computer and logged-in Chrome instances, form teams, run long tasks in the background, and switch to using Claude Code, Codex, or Pi as the runtime.

Why It Matters The proprietary MoE inference engine places frequently used experts on the GPU and the rest in memory or on disk, enabling larger models to run on smaller VRAM; it organizes emails and chats into a wiki nightly for agent retrieval.

Stars 1403 (+40 / 24h)

Reasons

  • 40 new stars in 24 hours, total stars 1403 (indicating attention only)
  • The README includes a generation speed table for seven MoE models on 8, 12, and 32 GB GPUs, noting that most measurements were simulated by limiting VRAM on larger GPUs

Usage

  • For macOS 14 or later on Apple Silicon, Windows x64, or Linux x64, download the installer from Releases, verify against SHA256SUMS.txt, and use Quick Install on first launch to set up services locally

Risks

  • The README promotes the tool using "Claude-level" claims and comparisons with third-party scores; the author claims 12 GB GPUs can run 125B models, but the first token takes about 3 minutes at this tier
  • The macOS package is unsigned and requires manually removing the quarantine attribute; agents have extensive permissions to operate logged-in browsers and desktop applications, and to read emails
  • The license is the Linger Open Source License, which adds conditions to Apache 2.0

Open

Trendshift(GitHub Trending 快照)

Explore the latest AI Radar updates