12GB GPU Runs a 125B Model: Linger Wants to Bring the Personal Agent Back to Local
Open source, local inference, zero token fees, and it can connect to Claude Code and Codex. Linger's pitch is straightforward, but whether a 16GB Mac actually runs smoothly is worth checking first, along with the official data and questions in the comments.
Linger is an open-source local Personal Agent from SimplexAI. The project was just released on GitHub, and its official pitch is direct: a Claude-level brain running on your computer, local inference with zero token fees, and the ability to connect to the official subscriptions of Claude Code and Codex.
The difference between a Personal Agent and a chat box is that it needs to read email, calendars, and Slack, remember projects and people, proactively remind you when necessary, and even use the browser and desktop apps to complete tasks. Cloud solutions have always had a problem: as context grows, token bills and privacy boundaries grow along with it.
Linger's approach is to keep the model, memory, and tool calls as local as possible. It defaults to the Pi framework, and a minimum of 16GB of Mac memory can run local models. The installer detects hardware and automatically selects a model. The official claim is that it can run 12 open-source models including Qwen, DeepSeek, and Gemma, with up to 5x inference acceleration from its in-house PolyStrata engine.
A more specific data point: a 125B Qwen3.8-Flash-Next runs at 45.5 tok/s on a 12GB RTX 3060 plus 64GB of memory. PolyStrata keeps the busiest experts of an MoE model on the GPU and puts the rest in memory or on an SSD, so model weights can exceed VRAM. This idea is not new, but Linger has built it into consumer-grade devices and a one-click installer.

Many figures in the official benchmark table come from simulated configurations, and real machines may be slower. For a 16GB Mac, Spark-X2.6-4B is recommended; with short prompts and two concurrent sessions, the slower path is about 33 to 37 tok/s, and after a long prompt of 57k tokens it is about 24 to 26 tok/s. Some commenters have asked whether a 16GB Mac can really run it smoothly, and there is not yet enough real-world feedback on that question.
There is also an opening for connecting subscriptions. Linger supports importing Claude Code, Codex, or the Pi runtime via CLI, and local models and subscription models can collaborate in the same room. One user mentioned that they pulled Codex and a local model into the same room for a Shenzhen site-selection recommendation; the interaction was smooth, and the final report UI was polished. This usage is more appealing to developers: expensive models make judgments, while local models handle repetitive work, which can bring the bill down a bit.
Memory is stored as a Wiki. Gmail, Outlook, Slack, Calendar, meetings, files, and agent conversations are archived into Wiki pages every night at 00:00, categorized as Daily report, Projects, Topics, People, and Companies. Officially, Linger scores 95% on the LoCoMo memory benchmark, slightly above Zep at 94.7% and Mem0 at 92.5%. A benchmark is a benchmark; whether it can remember in real work that a client changed their position last week is another matter. Still, keeping memory on a local device is valuable for people who are sensitive about privacy.

Another capability is proactive reminders. Another Me looks at the inbox, calendar, and Slack, and decides according to rules whether to speak now, follow up later, or stay quiet. Every night, Dreaming reads how the messages it sent were received, then rewrites its judgment rules without retraining the model. This design feels more like a personal assistant than simply piling up memory, and it also tests the boundaries of rules more.
Plugins connect through Composio to Gmail, Outlook, Google Calendar, GitHub, Notion, Slack, Discord, Figma, Canva, Higgsfield, Fal, and others, with 10,000 free calls per day. The original article also mentions that the Dmit IP gets locked and you need to work around it to register a Key. This detail shows that the pitfalls of early-stage tools are still there.
Privacy needs to be clear. Local models and Wiki memory can stay on your computer, but connecting to cloud models or voice still goes through external providers. The official FAQ states that voice requires Soniox and xAI. People who want to be fully offline should read the configuration carefully.
Installation is not complicated. GitHub has desktop packages: macOS 14+ on Apple silicon, Windows x64, and Linux x64 with glibc 2.35+. Intel Macs use Docker Compose. The official website's download section still says Coming soon, while GitHub already has a release. Those who want to try it should first accept the installation pitfalls that early-stage projects may have.
In the announcement, Tony Zhang cites the history of centralized computing moving to PCs in the 1960s and says intelligence should be distributed: your context, your data, and the intelligence that drives them should belong entirely to you. The analogy may not hold, but the cost curve of local inference and privacy needs are indeed pushing Personal Agents toward local devices. Today you still have to agonize over whether a 16GB Mac runs smoothly; tomorrow you may only need to agonize over electricity bills and GPUs. The GitHub link is here; before you Star it, look at the official benchmark conditions first.
发布时间: 2026-11-19 13:28