Skip to content

Linger runs a 125B model on a 12 GB GPU, fully local

simplexai-labs' open-source personal agent pairs a local MoE inference engine with nightly memory and a proactive assistant that rewrites its own rules.

Draftship Editorial4 min read

Linger, an open-source personal agent from simplexai-labs, had 1,622 stars and 126 forks on GitHub as of 2026-11-20, ten days after the repository was created on 2026-11-10. The pitch behind the Linger local AI agent is that a useful assistant should run on the machine in front of you rather than on a provider's servers. The code is mostly TypeScript, and the repository's topics read like a feature list: agent-teams, local-llm, memory, personal-agent, proactive-agent, proactive-ai. The stated goal is to reduce dependence on large providers by making capable models practical on ordinary computers.

What the agent actually does

The README describes one assistant that reads your inbox, calendar and Slack, files what it learns into wiki pages overnight, and messages you first when something needs attention. The project calls this half of the product Another Me. Each night a process named Dreaming reviews how those outgoing messages landed and rewrites the rules Another Me decides by, so a repeated refusal changes future behaviour without retraining the model. Naming the change you want, such as a different hour or a different kind of alert, can turn it into a rule in one pass.

The rest of the feature list, as the source describes it:

  • Desktop and browser control, where agents open apps on macOS, Windows or Linux and click and type. On Mac and Windows they work inside your everyday Chrome with the sign-ins you already have.
  • Agent teams, where one team.md file is shared by every agent on a job, with missing agents created on demand.
  • Task delegation, where long work moves to background copies of an agent while the conversation continues.
  • Model scheduling, where one local model is shared by all agents, approvals and background jobs, with work you are waiting on served first.
  • Bring your own runtime, running Claude Code, Codex or Pi on local models or on a subscription you already pay for.
  • Voice, with meeting recording, live transcript and follow-ups handed to Another Me.

Fitting a 125B model on a 12 GB card

The hardware claim is the reason the repo is being watched. The README says the 125B Qwen3.8-Flash-Next model runs on a 12 GB GPU with 64 GB of system RAM, using an in-house MoE inference engine called PolyStrata that keeps a model's busiest experts on the GPU and the rest in RAM or on an SSD. The same section claims zero token fees for local inference.

An installer called Model Autopilot reads your GPU, memory and bandwidth, keeps only configurations that fit with at least a 64K context, picks the smartest model and then the fastest, and counts a model as installed only after it answers a real question. The project says its lab search covered 49 language models, two inference engines, quantizations from 1-bit to 8-bit and context windows from 64K to 256K, tested tier by tier on NVIDIA cards and Macs.

The numbers come with footnotes

Most decode speeds in the published table were measured on simulated tiers, where a larger card is held to the tier's memory budget, so a real machine of that tier can be slower, the README states. The 38.1 tokens per second shown for a laptop RTX 4070 with 8 GB is flagged as not part of PolyStrata's results: it came from a Q2_0 file run on a 96 GB card capped to an 8 GB budget, with 32 GB of RAM and a 64K context. The 12 GB tier entry lists 21.3 tokens per second on short prompts and a first token after about three minutes. Comparisons against named cloud models use third-party AIHOT scores, where 50 is the reference average and scores within 3 points count as the same level.

Limits and open questions

The source leaves several things open. It does not state a licence. It does not report independent verification of the speed or quality numbers, and the 125B claim has not been shown on a real 12 GB card in public results. The 12 GB tier is marked as not that machine's default recommendation, so which model the installer picks there is unclear. On setup, the Docker Compose path ships with a default owner account and the test password test1234, which the README tells you to change before anyone else can reach the server, and the macOS package is unsigned, so the first launch is blocked until you clear the quarantine flag in Terminal.

What this suggests for building apps

Linger is a personal agent runtime, not an app builder. If the model and harness co-design approach holds up under outside testing, it likely lowers the floor for features that need a lot of inference, because local runs carry no per-token bill. It does not hand you a hosted web app with a database, sign-in and a domain, and the source never claims it does. For that side of the work, Draftship is an AI app builder where you describe an app in plain language and it writes the code, builds and hosts it, with a built-in database and user sign-in. Its guide to adding AI to your app covers wiring a model into something you ship.

You can read the full feature list, benchmark tables and install steps in the Linger repository.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.