Ex-DeepSeek researcher launches Linger: open-source agent runs a 125B model on 12 GB
Tony Zhang, who led the DeepSeek-VL project, has released Linger through his new start-up. It's an open-source PC agent that reads your email, calendar and meetings, keeps everything in a wiki and pings you on its own. It also claims a custom inference trick that runs a 125-billion-parameter model on a mid-range graphics card. Every performance number so far comes from the developers themselves.

What's going on
Tony Zhang (Zhang Bo in Chinese) worked on DeepSeek V1 and led DeepSeek-VL, the Chinese AI lab's vision-language model. He has now founded his own start-up, and its first product is Linger, a personal AI agent that runs on your own computer. An agent is an AI that doesn't just chat. It carries out tasks on its own: it operates programs, clicks through a browser and reads your email.
Linger is open source and lives on GitHub at simplexai-labs/linger. Version 0.1.0 came out on November 10, 2026, for macOS (version 14 and up, Apple Silicon only), Windows x64 and Linux x64. Tech blog explainx.ai reports that version 0.1.5 was already out by November 19 and that the project had gathered around 1,800 GitHub stars in ten days. The software is free, and if you use local models you pay no per-token fees.
Zhang explained the idea to Chinese tech outlet 36Kr by pointing to the history of the PC:
In the 1960s, computing power was centralized, until personal computers put it in everyone's hands. Today, AI models are trained in data centers and then rented out to you as tokens. We believe intelligence should be distributed.
The headline claim: 125 billion parameters on 12 GB
The most eye-catching claim is about the PolyStrata inference engine. A 125-billion-parameter model normally needs around 39 GB of VRAM even at 2-bit quantization, according to an analysis on note.com. (Quantization shrinks a model by storing its numbers at lower precision.) That won't even fit on an RTX 5090.
PolyStrata takes advantage of the fact that the supported model, Qwen3.8-Flash-Next, is a Mixture-of-Experts (MoE) model. It's made up of many specialized sub-networks ("experts"), and only a few of them are active for each token. PolyStrata keeps the frequently used "hot" experts in fast VRAM, moves less-used ones to system RAM and parks rarely used ones on the SSD. Think of a desk where the essentials sit within reach while everything else waits on the shelf or down in the basement.
Here's what the developers claim in the README:
| Hardware | Model | Speed |
|---|---|---|
| RTX 3060 (12 GB) + 64 GB RAM | Qwen3.8-Flash-Next 125B | 45.5 tokens/s |
| RTX 5090 (32 GB) | Qwen3.8-Flash-Next 125B | 179 tokens/s |
| various cards | other models | approx. 65 to 480 tokens/s |
There's also Model Autopilot. The installer detects your GPU, VRAM, system RAM and memory bandwidth. It then picks the "smartest model that fits this PC" from 49 models, 2 engines, quantization levels from 1 to 8 bits and context lengths from 64K to 256K. Before it finishes, it checks that the model can answer a test question. The models it can choose include Qwen3.8-27B, Qwen3.6-35B-A3B, Gemma 4 26B-A4B, LFM2.5-24B-A2B and Spark-X2.6-4B, which is the recommendation for 16 GB Macs.
explainx.ai points out caveats that the README itself admits. Many of the numbers come from simulated hardware tiers, and real machines may be slower. On the 12 GB tier, the first token after a 62,000-token prompt can take around three minutes to appear. That's a real problem for an agent that handles a lot of context. explainx.ai also notes that the 125B setup isn't even the default recommendation for 12 GB machines.
"Another Me": an agent that reaches out on its own
The centerpiece is an agent called "Another Me". It reads your inbox, calendar, chats and meeting notes locally and condenses them every night into a personal wiki of Markdown pages that all agents can search. Every change to its memory is logged along with the reason for it.
The agent decides for itself when to speak up. 36Kr gives an example: if a client emails on Monday asking for a quote by Friday, it won't remind you until Wednesday morning. A rescheduled meeting, on the other hand, gets flagged right away. A nightly process called "Dreaming" looks at how you reacted to its alerts and turns that into rules. One might be "supplier invoices → notify immediately". Another might keep it quiet about package notifications once you've ignored them twice. Linger calls this "implicit reinforcement", and it doesn't require retraining the model.
Other features:
- Desktop and browser control via Playwright and Cua. On Mac and Windows, the agent works inside the Chrome session you're already logged into, so you don't have to set up credentials again. If it gets stuck, you can take over.
- Agent teams: For big tasks, Another Me spins up specialized agents that follow a shared plan in a file called team.md.
- Prioritization: All agents share one local model. Whatever you're waiting on at the screen takes priority over background jobs.
- Bring your own runtime: Instead of local models, you can plug in Claude Code, Codex or Pi. Since version 0.1.4, you can also use your own API keys for OpenAI-compatible and Anthropic endpoints.
- Voice: Meetings are transcribed live with LiveKit and Soniox speech recognition, and you can ask questions about them afterward.
Numbers with question marks
Linger advertises a "Claude-class brain". To back that up, the project uses its own "SimplexAI Score", where 50 equals the average of 123 reference models. The sources report conflicting numbers here. According to note.com, a setup with a 12 GB GPU and 32 GB of RAM scores 54.8 (beating Claude Sonnet 4.6 at 53.6), and one with a 24 GB GPU and 64 GB of RAM scores 59.3. 36Kr, however, gives 57.3 for 12 GB + 32 GB of RAM, "roughly Claude Opus 4.6", and 51.9 for a 24 GB Mac. The figures apparently changed between versions. explainx.ai notes that the comparison models are no longer state of the art, and that a blended score hides which tasks a local model fails at.
On LoCoMo, a benchmark for long-term memory in conversations, note.com and 36Kr both cite 95.0 percent. That's ahead of Zep (94.7), Mem0 (92.5), Letta (74.0) and OpenAI Memory (52.9). explainx.ai, however, says the current README doesn't publish a memory benchmark of its own and instead points to the open-source project Hindsight for fact extraction and search. All sources agree that there are no independent tests yet.
Tech and security
According to note.com, the codebase is just under 95 percent TypeScript and Go. A "Computer Daemon" runs on your PC and does the actual work, while persistent data lives on a "TeamHub" server backed by PostgreSQL. The daemon opens no inbound ports and only holds an outbound WebSocket connection. In the workspace, only playground/ is writable. memory/ and llm-wiki/ are read-only. Clicks, keystrokes and shell commands pass through a two-stage guardrail: a rule check first, then a review by a language model. The one exception is the Pi runtime, which currently skips the check.
What you need to try it
The desktop app comes as a .dmg for Apple Silicon Macs (macOS 14+), as an installer or portable .zip for Windows and as a .tar.gz for Linux (glibc 2.35+). Intel Macs are only supported through Docker, and the Docker VM needs at least 5 GiB of memory. For local models, the developers say a 16 GB Mac is enough for the small Spark-X2.6-4B. For the 125B model, they list a 12 GB Nvidia card like the RTX 3060 plus 64 GB of RAM and a fast SSD. If your hardware doesn't fit, you can use your own API keys.
Gotchas: The macOS app isn't notarized by Apple, so on first launch you have to strip the quarantine attribute with xattr. explainx.ai advises doing that only after you've checked the file against the included SHA256SUMS.txt. The Windows installer is unsigned and triggers a SmartScreen warning. If you go the Docker route, change the test password right away. There's also a known bug where local services keep running after you quit from the taskbar.
The bigger picture
Linger isn't a new language model. It's a bundle of inference engine, model selection, memory and agent infrastructure, all tuned to work together. The bet: if you optimize the model and the agent software together, an affordable graphics card can do what would otherwise take data center hardware, with no token costs and no data leaving your machine. That sets Linger apart from agents like Hermes, which usually assume a hosted model.
There's plenty to be skeptical about. All performance claims come from the developers, much of it was measured on simulated hardware, and the numbers differ between sources. Offloading experts to the SSD gets slow when the agent keeps switching between email, code and meetings. The license isn't plain Apache 2.0 and adds extra conditions, so if you want to use it commercially, read LICENSE and NOTICE first. And an agent that controls your inbox, calendar, Slack and logged-in browser, and acts on its own initiative, has very broad access. explainx.ai recommends starting with a separate browser profile and restricted accounts. At version 0.1.x with dozens of open pull requests, you should also expect frequent changes, some of them breaking. Still, the project is worth watching, above all as a blueprint for what a fully local personal assistant could look like.
Sources
- Bing News (ko-KR) minsimnews.com
- note.com note.com
- explainx.ai explainx.ai
- eu.36kr.com eu.36kr.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.