Linger: The Open-Source Personal Agent That Runs a 125B Model on Your PC Simplexai-labs released Linger, an open-source personal agent that runs locally on a PC or Mac, and the project's GitHub repository gained about 1,800 stars in the ten days after its first release on November 10, 2026. Linger's README claims its PolyStrata mixture-of-experts runtime runs a 125-billion-parameter Qwen3.8-Flash-Next model on a 12 GB GPU plus 64 GB of system RAM at 45.5 tokens per second on an RTX 3060 and 179 tokens per second on an RTX 5090; the project has not been independently tested, and the README states many figures come from simulated hardware tiers. The latest release, Desktop v0.1.5, was published November 19 under the Linger Open Source License, an Apache 2.0 derivative, and is free to download with no token fees for local inference. Linger is a new open-source personal agent that runs on your own PC, and it gained about 1,800 GitHub stars in the ten days after its first release on November 10, 2026. The simplexai-labs/linger repository https://github.com/simplexai-labs/linger?ref=explainx pitches "a Claude-class brain running locally": a 125-billion-parameter mixture-of-experts model on a 12 GB GPU with 64 GB of system RAM, plus memory, a proactive assistant and desktop control. The speed and quality figures come from the project's own README. We have not seen independent tests yet, so treat them as claims. This post explains what Linger is, how its pieces fit together, how to install it, and where the README is careful about its own limits. Linger at a glance | Question | Answer from the README and releases page | |---|---| | What is it? | An open-source personal agent that runs locally on a PC or Mac | | Who builds it? | simplexai-labs on GitHub | | Latest release | Desktop v0.1.5, published November 19 | | Platforms | macOS 14+ on Apple silicon, Windows x64, Linux x64 glibc 2.35+ | | Headline claim | 125B MoE model Qwen3.8-Flash-Next on a 12 GB GPU plus 64 GB RAM | | Reported speed | 45.5 tokens per second on an RTX 3060 12 GB; 179 on an RTX 5090 | | License | Linger Open Source License, based on Apache 2.0 with extra conditions | | Cost | Free to download; no token fees for local inference | What does Linger actually do? The README describes five layers that work together. A local model runtime called PolyStrata. It is a mixture-of-experts engine that keeps the most-used experts on the GPU, moves less-used ones into system RAM and leaves rarely used ones on the SSD. The idea is the same one behind other consumer-hardware MoE tricks we have covered, such as InferSSD running very large models through llama.cpp https://explainx.ai/blog/inferssd-800b-consumer-hardware-llama-cpp-november-2026 . MoE models only activate a few experts per token, so the hot ones can live in fast memory while the cold ones wait on slower storage. Model Autopilot. The installer detects your GPU, VRAM, RAM and memory bandwidth, picks the smartest model that fits with at least a 64K context window, and checks that it can answer a test question before it finishes. A wiki and memory layer. Mail, chats and meetings are turned into wiki pages that update nightly. Agents search those pages instead of rereading raw history. The README credits the open-source Hindsight project for fact extraction and search, and publishes no memory benchmark of its own. "Another Me", a proactive agent. It watches your inbox, calendar and Slack, follows up on its own, and messages you only when needed. A nightly process the README calls Dreaming rewrites its decision rules based on how you reacted. It can also assemble agent teams that share a plan file named team.md . Desktop and browser control. Agents can operate desktop apps and, on Mac and Windows, work inside your signed-in Chrome. Long jobs run as background copies of an agent so the chat stays responsive, and all agents share one local model with priority for user-facing work. Voice is part of the package too: meeting recording with live transcripts and a handoff to follow-up tasks. How does the "125B on a 12 GB GPU" claim work? Total parameters and active parameters are different numbers. A mixture-of-experts model can have 125 billion parameters on disk while using only a small slice for each token. PolyStrata exploits that by tiering the experts across GPU, RAM and SSD. The README's benchmark table lists decode speed for several models. For Qwen3.8-Flash-Next it gives 45.5 tokens per second on an RTX 3060 12 GB and 179 on an RTX 5090 32 GB . Other listed models run from roughly 65 to 480 tokens per second depending on the card. The README is open about limits. It says many numbers come from simulated hardware tiers, so real machines may be slower, and one 8 GB result was measured under simulated conditions. It also notes that on the 12 GB tier the first token can take about three minutes after a 62K-token prompt. That matters for an agent that stuffs a lot of context into every call. Read the numbers as a ceiling you may hit after warm-up, not a promise for every prompt. If you want the general trade-offs of running big models locally, our guide to llama.cpp 1.0 and what changes https://explainx.ai/blog/llama-cpp-v1-0-0-stable-libllama-libggml-semantic-versioning-what-changes-2026 is a good companion. Which models does it support? The benchmark tables name these local models: - Qwen3.8-Flash-Next 125B, with a Q2 0 variant noted - Qwen3.8-27B - Qwen3.6-35B-A3B - LFM2.5-24B-A2B - Flint 40B-A2.5B - Ornith 1.5 - Gemma 4 26B-A4B - Tiel Coder 35B-A3B - Spark-X2.6-4B, recommended for Macs with 16 GB Model Autopilot picks among them based on your hardware. If you would rather not run a local model at all, release 0.1.4 added bring-your-own-key support for OpenAI-compatible and Anthropic endpoints. The README also lists Claude Code, Codex and Pi as runtimes you can plug in, either on local models or on subscriptions you already pay for. How good is the "Claude-class" claim? The README scores recommended models on what it calls the SimplexAI score, built on a third-party AIHOT scale where 50 is the reference average and scores within 3 points count as the same level. Local recommendations range from 35.4 to 57.3 across hardware tiers. The README maps them to cloud models: 51.9 is "about Claude Opus 4.5", 54.0 is "about Claude Sonnet 4.6", and 57.3 is "about Claude Opus 4.6". Two cautions apply. First, those comparison models are older than the current frontier, which now includes models such as the one in our Gemini 4 Neon launch post https://explainx.ai/blog/gemini-4-neon-launch-price-benchmarks-vs-carbon-hydrogen-november-2026 . Second, the 57.3 row is the Qwen3.8-Flash-Next setup, which the docs say is not the default recommendation for a 12 GB machine and was measured on a simulated tier. A composite score also hides which tasks a local model fails. "Claude-class" here means close to a specific older Claude tier on one blended score. How do I install Linger? There are two routes. Desktop app recommended for one person . 1. Download the package for your platform from the releases page https://github.com/simplexai-labs/linger/releases?ref=explainx : a .dmg for macOS, an .exe or portable .zip for Windows, or a .tar.gz for Linux. 2. Verify the file against SHA256SUMS.txt , which is attached to each release. 3. On macOS, the app is not notarized. The README points to a first-launch guide that removes the quarantine attribute with xattr . That bypasses a macOS security check, so do it only after the checksum matches. 4. Windows shows a SmartScreen warning on first run because the installer is unsigned. Docker Compose shared servers or building from source . 1. Clone the repo and copy .env.example to .env . 2. Run docker compose up --build --detach --wait . The Docker VM needs at least 5 GiB of memory. 3. Create an owner account with the teamhub-account command. 4. Open http://localhost:3000 and sign in. The README's Docker example uses a test password. Change it under Settings, then Security, before anyone else can reach the server. For local models, run ./scripts/install/install.sh on macOS or Linux, or install.ps1 on Windows. The installer checks hardware, installs a model after you approve it, and connects the computer to the server. There is no Intel Mac package, so Intel Macs must use Docker. Release history: a young project moving fast The releases page shows how quickly this is changing: | Version | Date | Notable changes | |---|---|---| | 0.1.0 | Nov 10 | First release for macOS Apple silicon , Windows and Linux | | 0.1.1, 0.1.2 | Nov 13, Nov 16 | Minimal release notes | | 0.1.3 | Nov 18 | Signed automatic updates, Local models settings page, Send Diagnostics, bundled computer-use tools | | 0.1.4 | Nov 19 | Bring-your-own-key endpoints, automatic Pi setup, chat prioritized over background work, guided Windows installer | | 0.1.5 | Nov 19 | Cross-room handoffs, routines at 10-minute intervals, updated local model catalog | The repository shows 541 commits, 8 open issues and 33 open pull requests at the time of writing. Known issues in 0.1.4 and 0.1.5 include the unsigned Windows installer, the un-notarized macOS app, and the fact that quitting from the tray leaves some local services running. What should you worry about? Permissions. An agent that watches your mail, calendar and Slack, and drives your signed-in browser, has broad access. A proactive agent acts without being asked, which is the feature and the risk. Start with a separate browser profile and limited accounts, and review what it does before you widen access. Our coverage of Claude Code Projects and its proactive coordinator https://explainx.ai/blog/claude-code-projects-proactive-coordinator-november-2026 covers the same design tension in a hosted product. License. "Apache 2.0 with additional conditions" is not plain Apache 2.0. The README points to a NOTICE file that lists third-party code and says some logos and images are not covered by the license. If you plan to build a product on it, read the license text first. Self-reported numbers. The speed table and the SimplexAI scores come from the project's own lab, and the docs say most were measured on simulated hardware tiers. We found no independent tests. Supply chain. A fast-growing repo with 33 open pull requests and frequent releases deserves normal care: pin versions, verify checksums, and read the update behavior. Since 0.1.3 the app checks for updates automatically and uses signed updates. How does it compare with other local and open agents? Linger sits next to open personal agents such as Nous Research's Hermes line, which we covered when it moved toward phones in Hermes mobile beta https://explainx.ai/blog/hermes-mobile-beta-nous-research-ios-android-full-after-thanksgiving-release-claimed-vs-verified-2026 , and the harness reliability work discussed in the Hermes Agent leaderboard analysis https://explainx.ai/blog/joel-niklaus-harness-leaderboard-reliability-hermes-agent-2026 . The difference is the bundle. Most agent harnesses assume a hosted model. Linger ships its own runtime, model picker, memory wiki and server together, and bets that tight co-design lets a modest GPU do work that usually needs a data-center card. That bet has a cost. Tiering experts onto SSD is slow when the working set changes quickly, which is exactly what happens when an agent switches between mail, code and meetings. A hosted frontier model will still be faster and stronger on hard tasks. Linger's pitch is privacy and zero token bills, not raw capability. Should you try it? Try it if you have a 12 GB or larger NVIDIA GPU with 64 GB of RAM, or a recent Apple silicon Mac, and you want a private assistant to experiment with. Start in a sandboxed account, check the checksum, and measure your own tokens per second on your own prompts before you trust any table. Wait if you need signed installers, a stable license position or independent benchmarks. At version 0.1.5, with a release roughly every day or two, expect rough edges. The stars show strong interest in local agents. They do not show that the claims hold on your machine. Related reading Version numbers, star counts and benchmark figures are accurate as of November 20, 2026 and come from the project's own README and releases page.