Article · AI translation
Built specifically for local deployment
Together, they form a blaze; apart, they shine like stars across the sky.
Tony Zhang, a former author of DeepSeek V1 and project lead of DeepSeek-VL, has just officially unveiled his new startup project: Linger.
Its focus is the currently popular personal agent.
X post video.mp4
Its approach is interesting, too: while peers are building apps, the team has chosen to prioritize computers.
That means Linger can operate the webpages, email inbox and calendar already on your computer, without the hassle of setting up a pile of login credentials.
If it gets stuck, you can also take over directly.
X comment video.mp4
Another point: as an intermediary layer, Linger supports connections to various third-party models.
It is also specifically optimized for local LLMs.
Specifically, an RTX 3060 graphics card with 12GB of VRAM can run the 125B-parameter Qwen3.8-Flash-Next at a generation speed of 45.5 tokens per second.
Oh, and there's another fairly important point—
The system is open source...
Why?
In his announcement post, Tony Zhang cited the history of the PC's rise:
In the 1960s, computing was centralized, until the PC put power in everyone's hands. Today, AI is trained in data centers and then rented to you in the form of tokens.
We believe intelligence should be distributed. Your context, your data, and the intelligence that powers them should all belong entirely to you.
In the 1960s, computing was centralized, until the PC put power in everyone's hands. Today, AI is trained in data centers and then rented to you in the form of tokens.
We believe intelligence should be distributed. Your context, your data, and the intelligence that powers them should all belong entirely to you.
So today, they open-sourced Linger, a personal agent that lives on your computer.
It carries on Big Fat Whale's fine tradition, too 😭.
An AI that takes initiative and learns on its own
At Linger's core is an agent system called "Another Me."
First and foremost, it is a memory hub.
Linger reads your inbox, calendar, chats and meetings locally, automatically organizing that information every night into a personal knowledge base (Wiki) that all agents can query.
On the LoCoMo benchmark, which measures long-term conversational memory, Linger achieved 95% accuracy, surpassing several memory-enhanced AI systems, including Mem0 and MemOS.
With all this, Linger can do a better job of working proactively.
It quietly watches your information stream and decides when to alert you based on an event's importance and urgency. If a client emails on Monday asking for a quote by Friday, it will wait until Wednesday morning to pop up a reminder instead of interrupting you right away.
But if a meeting time changes, it will notify you immediately.
The agent also learns from your feedback, which Linger calls "implicit reinforcement."
For example, when it reminds you that an invoice is nearly due and you reply, "It arrived last week; you should have told me then," the system notes that as a clue.
If a similar situation arises with another invoice and you express the same preference again, the system automatically creates a rule: "Supplier invoices → alert immediately."
Conversely, if you show no interest in "delivery alerts" twice in a row (for example, by ignoring them or replying, "Don't tell me"), it will also learn to stay quiet in the future.
You can also teach it in one go with a direct instruction. For example: "Notify me immediately when a client replies," or "After 9 p.m., you don't need to tell me right away about things like meeting changes."
This ongoing interaction and learning helps the agent understand your work habits better and better, becoming a truly personalized assistant.
Beyond that, Linger can also:
Operate the desktop and browser.
As mentioned at the start, Linger can directly use Chrome and various apps that are already logged in on your computer, without reconfiguration.
Assemble a team of agents.
This is the Super Agent principle that DeepSeek explained previously.
When you give it a complex task, Another Me calls on or creates agents with different roles to work together to complete it.
Meanwhile, you only need to stay in your single session with Linger.
X comment video.mp4
Voice interaction.
This is mainly intended for office work.
Linger can transcribe meetings in real time; afterward, you can ask the agent about the meeting and have it follow up on action items.
In short, it is a personal agent optimized specifically for office work.
However, some commenters did not quite understand the product's positioning:
Why not develop it on the DeepSeek harness? 🤣 How is it different from ds-bot?
Why not develop it on the DeepSeek harness? 🤣 How is it different from ds-bot?
Tony Zhang responded:
Linger focuses on optimizing the model and harness together: a 12GB graphics card can run a 125B model locally, with no per-token fees and no data leaving the machine.
In terms of product positioning, DS-Bot centers on collaboration between bots and between bots and groups; Linger centers on people and their own agents, and on agents and the local environment.
Linger focuses on optimizing the model and harness together: a 12GB graphics card can run a 125B model locally, with no per-token fees and no data leaving the machine.
In terms of product positioning, DS-Bot centers on collaboration between bots and between bots and groups; Linger centers on people and their own agents, and on agents and the local environment.
Running a 125B model on a 12GB graphics card
The core technology that lets Linger run a model with over 100 billion parameters on consumer hardware is a MoE inference engine called PolyStrata.
Its principle can be understood as an intelligent tiered storage mechanism.
For a large MoE (mixture-of-experts) model, PolyStrata puts the most frequently used and active "expert modules" in the fastest GPU memory, less frequently used ones in system memory, and the least frequently used ones on an SSD.
When the model needs to call on a particular expert, the system dynamically moves data among the GPU, system memory and SSD. This means the entire model, which can be tens or hundreds of GB in size, does not have to fit into limited GPU memory.
The results are immediate.
According to official figures, a 125B-parameter Qwen3.8-Flash-Next model can generate 45.5 tokens per second on a computer with a 12GB RTX 3060 and 64GB of system memory.
On the more powerful RTX 5090 (32GB), the speed jumps to 179 tokens per second.
To make local large models easier to use, Linger has also built a "Model Autopilot" feature.
It automatically detects your hardware configuration (GPU, VRAM and system memory), then recommends and installs the best-performing model for you from combinations of 49 language models, 2 inference engines, multiple quantization options and different context lengths.
According to SimplexAI's scoring system, local devices with different configurations can achieve capabilities comparable to leading cloud-based large models:
8GB graphics card + 32GB system memory: score 35.4, roughly equivalent to Gemini 2.5 Pro.
Mac with 24GB of unified memory: score 51.9, roughly equivalent to Claude Opus 4.5.
12GB graphics card + 32GB system memory: score 57.3, roughly equivalent to Claude Opus 4.6.
This means your gaming laptop or MacBook will have a genuinely powerful AI brain available 24/7, with no token fees and with all data kept local.
DeepSeek-VL lead starts another venture
Behind Linger is a team with deep technical expertise, and its founder Tony Zhang has a particularly impressive résumé.
(It is a DeepSeek team, after all...)
Tony Zhang earned his undergraduate degree from Zhejiang University and his PhD from the Hong Kong University of Science and Technology in 2019. He then joined Microsoft Research Asia (MSRA), where he spent years researching computer vision and generative models, and was an author of several well-known projects, including VQ-Diffusion and Rodin 3D digital humans.
In 2023, he joined DeepSeek as an AI scientist.
There, he worked on the early DeepSeek LLM V1 series of language models and served as project lead for the vision-language model DeepSeek-VL. The model has been widely praised for its ability to understand images, documents and charts in real-world settings.
After leaving DeepSeek, Tony Zhang joined the faculty at Zhejiang University as a Hundred Talents Program researcher and doctoral supervisor, and founded Simplex AI.
Linger is Simplex AI's latest creation.
And the inference engine at Linger's technical core, built specifically for local deployment?
According to Tony Zhang, it was built by an 18-year-old prodigy on the team.
18 years old? Just graduated from high school?