Newshot

DeepSeek V1 Author's Startup Project Revealed: Open-Source Agent

AI summary

Built specifically for local deployment.

Translation(model translation, not the source's own text)

Just Now, DeepSeek V1 Author's Startup Project Revealed: Open-Source Agent

Gathered, we are a fire; scattered, we are stars across the sky.

Just now, Tony Zhang, former DeepSeek V1 author and Project Lead of DeepSeek-VL, officially announced his new startup project: Linger.

The direction is precisely the hottest thing right now: the personal Personal Agent.

The approach is also quite interesting. While peers are all building apps, they prioritized the computer.

This means Linger can operate the web pages, inbox, and calendar already on your computer, with no need to fiddle with a pile of login credentials.

When it gets stuck somewhere, you can also take over directly yourself.

Another point: as a middle layer, Linger supports connecting to various third-party models.

And it has been specifically optimized for local LLMs.

Specifically, an RTX 3060 graphics card with 12GB of VRAM can run a 125B-parameter Qwen3.8-Flash-Next, at a generation speed of 45.5 tokens/second.

Oh right, there's also a rather important point—

In the announcement post, Tony Zhang cited the history of the PC's rise:

In the 1960s, computing was centralized, until the PC put power into 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 drives them should all be entirely owned by you.

In the 1960s, computing was centralized, until the PC put power into 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 drives them should all be entirely owned by you.

So today, they open-sourced Linger, an exclusive Agent that lives on your personal computer.

This also continues the fine tradition of the big fat whale 😭.

An AI That "Takes Initiative and Can Self-Learn"

At the core of Linger is an Agent system called "Another Me."

Linger locally reads your inbox, calendar, chat records, and meetings, and every night automatically organizes this information into a personal knowledge base (Wiki) for all Agents to query.

On the LoCoMo benchmark, which measures long-term conversational memory capability, Linger's accuracy reached 95%, surpassing multiple memory-enhanced AI systems such as Mem0 and MemOS.

With these, Linger can work proactively more effectively.

It silently observes your information stream and decides when to remind you based on the importance and urgency of events. If it receives a client email on Monday requesting a quote by Friday, it won't disturb you immediately, but will wait until Wednesday morning to pop up a reminder.

But if it's a meeting time change, it will notify you immediately.

This Agent also self-learns from your feedback, which Linger calls "implicit reinforcement."

For example, when it reminds you that an invoice is about to expire, and you reply, "I received it last week, you should have told me then." The system will record a clue.

If the next time a similar situation arises with another invoice and you again express the same preference, the system will automatically generate a rule: "Supplier invoice → remind immediately."

Conversely, if you show disinterest in "delivery reminders" twice in a row (for example, by ignoring them directly or replying "don't tell me"), it will also learn to stay silent in the future.

You can also give direct instructions to teach it all at once. For example, tell it: "Notify me immediately about client replies," or "After 9 PM, no need to immediately mention things like meeting changes."

This continuous interaction and learning makes the Agent increasingly understand your work habits, becoming a truly personalized assistant.

In addition, Linger can also:

As mentioned at the beginning, Linger can directly use Chrome and various Apps already logged in on your computer, with no need to reconfigure.

This is the Super Agent principle that DeepSeek explained before.

When you give a complex task, Another Me will invoke or create Agents with different functions to collaborate on the work.

And you only need to stay in this one Session with Linger.

This point mainly serves office scenarios.

Linger can transcribe meetings in real time, and after the meeting directly ask the Agent about the meeting content and have it follow up on to-dos.

In short, it is a Personal Agent specially optimized for office scenarios.

However, some netizens don't quite understand this product positioning:

Why not develop based on the DeepSeek harness 🤣 What's the difference from ds-bot?

In response, Tony Zhang said:

Linger's focus is optimizing the model and harness together: a 12GB graphics card can run a 125B model locally, with no per-token payment, and data never leaves the machine. From a product positioning perspective: the core of DS-Bot is Bot-to-Bot and Bot-to-group collaboration; the core of Linger is the collaboration between a person and their own Agent, and between the Agent and the local environment.

Linger's focus is optimizing the model and harness together: a 12GB graphics card can run a 125B model locally, with no per-token payment, and data never leaves the machine.

From a product positioning perspective: the core of DS-Bot is Bot-to-Bot and Bot-to-group collaboration; the core of Linger is the collaboration between a person and their own Agent, and between the Agent and the local environment.

A 12GB Graphics Card Running a 125B Model

To let consumer-grade hardware run hundred-billion-parameter large models, Linger's core technology is a MoE inference engine called PolyStrata.

Its principle can be understood as an intelligent tiered storage mechanism.

For a massive MoE (Mixture of Experts) model, PolyStrata places the most commonly used and most active "expert modules" in the fastest GPU VRAM, places the less commonly used ones in memory, and places the least frequently used ones on the SSD solid-state drive.

When the model needs to invoke a certain expert, the system dynamically moves data among the GPU, memory, and SSD. In this way, there is no need to stuff the entire tens or hundreds of GB model into limited VRAM.

According to official data, a 125B-parameter Qwen3.8-Flash-Next model can achieve a generation speed of 45.5 tokens/second on a computer equipped with a 12GB RTX 3060 and 64GB of memory.

On a stronger RTX 5090 (32GB), the speed soars to 179 tokens/second.

To make it more convenient for users to use local large models, Linger also created a "Model Autopilot" feature.

It automatically detects your hardware configuration (GPU, VRAM, memory), and then from combinations of 49 language models, two inference engines, multiple quantization schemes, and different context lengths, recommends and installs the most powerful model for you.

According to SimplexAI's scoring system, local devices with different configurations can achieve capabilities comparable to mainstream cloud large models:

  • 8GB graphics card + 32GB memory: score 35.4, roughly equivalent to Gemini 2.5 Pro.
  • Mac 24GB unified memory: score 51.9, roughly equivalent to Claude Opus 4.5.
  • 12GB graphics card + 32GB memory: score 57.3, roughly equivalent to Claude Opus 4.6.

This means your gaming laptop or MacBook will have a truly powerful AI brain that is on standby 24/7, has no token fees, and keeps data entirely local.

DeepSeek-VL Lead Starts Up Again

Behind Linger is a team with a deep technical background, and the resume of its founder Tony Zhang is particularly impressive.

(After all, it's from the DeepSeek lineage…)

Tony Zhang graduated from Zhejiang University for his undergraduate degree, obtained his PhD from the Hong Kong University of Science and Technology in 2019, and then joined Microsoft Research Asia (MSRA), where he long engaged in computer vision and generative model research and is the author of multiple well-known projects such as VQ-Diffusion and Rodin 3D digital humans.

In 2023, he joined DeepSeek as an AI scientist.

Here, he not only participated in the R&D of the early language model DeepSeek LLM V1 series, but was also the Project Lead of the vision-language large model DeepSeek-VL. This model has been highly praised for its ability to understand images, documents, and charts in real-world scenarios.

After leaving DeepSeek, Tony Zhang entered Zhejiang University to teach, serving as a Hundred Talents Program researcher and doctoral supervisor, and founded Simplex AI.

Linger is the latest work of Simplex AI.

And as the technical core of Linger, that inference engine built specifically for local deployment?

According to Tony Zhang, it was built by an 18-year-old prodigy on the team.

  • Nvidia and AMD Join Forces for the First Time to Invest! Genesis Completes Series A Funding at a Valuation Exceeding 23 Billion 2026-11-18
  • Multimodal "Source God"! MiniMax Open-Sources SOTA Image Generation Model, Native 4K, Two Dimes per Image 2026-11-12
  • Robots Take Over Hotel Laundry! Dual-Brain Collaboration + RSI, Top-Tier Astra Driver Arrives 2026-11-11
  • "AI Ruined My Computer Science Degree" 2026-11-09

Related Reading

DingTalk Is Truly Bold! First to Integrate DeepSeek, the First in the Alibaba Ecosystem to Eat the Crab

Liang Wenfeng Contributes 20 Billion! DeepSeek's Record-Breaking First Funding Round of 50 Billion, V4.1 Slated for June

If it ultimately materializes, this will be the largest funding round in the history of Chinese large model companies

Exposed for Distilling DeepSeek and Faking It! The European Version of OpenAI Collapses

Distilled models must be labeled; transparency is the key.

Deepseek's New Model Accidentally Exposed! Coding Benchmark Scores Surpass Claude 3.5 Sonnet in One Go

It also ascended to the throne of the strongest open-source LLM on LiveBench

"DeepSeek Even Bypassed CUDA" — Paper Details Spark Renewed Discussion, Engineer's Soulful Question: Is Nvidia's Moat Still There?

Nvidia has just recovered from the 4-trillion-yuan plunge triggered by DeepSeek-R1, and is now facing new pressure?

DeepSeek's Low-Price Storm Subdues Silicon Valley! Overseas Platforms Compete to Subsidize V4 Flash

What else is there to say, Sage Liang — I'll just give you all my subscription fee

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 量子位
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗