A 125B model on a 12 GB GPU, and it remembers your inbox
Linger is an open-source personal agent that runs a large local model on ordinary Macs and PCs by shuffling MoE experts between GPU, RAM and SSD.
What it does Linger is a local-first personal agent: it connects your mail, chats, calendar and meetings, builds a shared memory from them, and can act on your behalf — assembling agent teams, using the browser, and asking for approval before it sends an email. The pitch is a “Claude-class brain” with zero token fees, since inference happens entirely on your machine. It’s built in TypeScript and powered by their own MoE inference engine, PolyStrata.
The interesting bit The trick is expert offloading: for each token, a router picks the few active MoE experts and keeps the hottest ones on the GPU, the lukewarm ones in RAM, and the cold ones on SSD. That’s how a 39 GB 2-bit quantization of a 125B model squeezes onto a 12 GB RTX 3060 at 45.5 tok/s. There’s also a “Model Autopilot” that benchmarks 49 models across quantizations and context lengths in their lab, then picks the smartest one that actually fits your hardware with at least 64K context.
Key highlights
- 125B Qwen3.8-Flash-Next runs on a 12 GB GPU with 64 GB RAM via PolyStrata expert tiering
- Nightly wiki/memory pipeline files mail, chats and meetings into a searchable per-entity graph — 95% on the LoCoMo memory benchmark, ahead of Zep, Mem0 and OpenAI Memory
- Model Autopilot detects your hardware and installs the best-fitting model, verified only after it answers a real question
- Per-tier recommendations: a 24 GB Mac lands roughly at Claude Sonnet 4.6-class on their SimplexAI score
- Runs fully offline; data stays on the device
Caveats
- Several benchmark numbers come from a simulated tier (a bigger card held to a smaller memory budget), so real machines may be slower — the README says so itself
- The 8 GB Flash-Next figure is a lab measurement with unsettled earlier numbers, and one 12 GB recommendation has a ~3-minute first token
- “Claude-class” rests on the project’s own SimplexAI scoring, not purely third-party evaluation
Verdict Worth a look if you want a genuinely local agent with real memory and don’t mind a young project with self-reported benchmarks. If you just want a chat UI over Ollama, this is far more opinionated — and far more ambitious.
Frequently asked
- What is simplexai-labs/linger?
- Linger is an open-source personal agent that runs a large local model on ordinary Macs and PCs by shuffling MoE experts between GPU, RAM and SSD.
- Is linger open source?
- Yes — simplexai-labs/linger is an open-source project tracked on heatdrop.
- What language is linger written in?
- simplexai-labs/linger is primarily written in TypeScript.
- How popular is linger?
- simplexai-labs/linger has 3.4k stars on GitHub.
- Where can I find linger?
- simplexai-labs/linger is on GitHub at https://github.com/simplexai-labs/linger.