type
Post
status
Published
date
Nov 21, 2026
slug
summary
tags
category
行业动态
icon
password
From DeepSeek-VL Lead to Open-Source Agent: Tony Zhang Launches Linger for Local AI
A new wave of personal AI agents is emerging that prioritizes local control over cloud dependency. Tony Zhang, the former project lead of DeepSeek-VL and a key contributor to DeepSeek LLM V1, has officially revealed his latest venture: Linger. Developed by his startup Simplex AI, this open-source project positions itself as a Personal Agent designed specifically for local deployment on personal computers. Unlike competitors focusing on mobile applications, Linger prioritizes the desktop environment, allowing users to leverage existing logged-in web sessions, email inboxes, and calendars without re-entering credentials or exposing sensitive login data.
An "Another Me" Agent with Proactive Learning
The core of Linger is a system named "Another Me," which functions as a memory hub. It locally reads user data from inboxes, calendars, chat records, and meetings, automatically organizing this information into a personal knowledge base (Wiki) each night. In long-term conversational memory benchmarks, specifically the LoCoMo test, Linger achieved an accuracy rate of 95%, outperforming other memory-enhanced systems like Mem0 and MemOS.
Beyond passive storage, the agent operates proactively. It monitors the user's information stream and determines when to issue alerts based on urgency and importance. For example, if a client email requests a quote by Friday, Linger might wait until Wednesday to remind the user, avoiding unnecessary interruptions. However, immediate notification is triggered for critical changes, such as meeting time shifts. The system utilizes "implicit reinforcement" to learn from user feedback. If a user repeatedly ignores shipping updates, the agent learns to remain silent on that topic. Conversely, if a user expresses a preference for immediate alerts on supplier invoices, the system generates a specific rule: "Supplier Invoice → Immediate Reminder." Users can also directly instruct the agent, such as specifying that client replies require instant notification or that meeting changes after 9 PM do not. This continuous interaction allows the agent to evolve into a truly personalized assistant that understands specific work habits.
Running 125B Models on Consumer Hardware
A defining technical feature of Linger is its ability to run large local language models on standard consumer hardware. The project introduced an inference engine called PolyStrata, a Mixture-of-Experts (MoE) mechanism that utilizes hierarchical storage. PolyStrata places the most frequently active "expert modules" in GPU memory, less used ones in system RAM, and the least frequent in SSD storage. The system dynamically moves data between these layers as needed, eliminating the requirement to load entire models into limited VRAM.
According to official data, a 125B parameter model, identified as Qwen3.8-Flash-Next, can achieve a generation speed of 45.5 tokens/second on a computer equipped with a 12GB RTX 3060 GPU and 64GB of memory. For higher-end setups with an RTX 5090 (32GB), the speed increases to 179 tokens/second. To simplify deployment, Linger includes a "Model Autopilot" feature that detects hardware specifications and recommends the optimal model from a pool of 49 options. Benchmark comparisons indicate that an 8GB GPU with 32GB RAM can achieve performance comparable to Gemini 2.5 Pro, while a Mac with 24GB unified memory rivals Claude Opus 4.5.
Tony Zhang's Background and Strategic Positioning
Tony Zhang holds a bachelor's degree from Zhejiang University and a Ph.D. from the Hong Kong University of Science and Technology. He previously worked at Microsoft Research Asia (MSRA), contributing to projects like VQ-Diffusion and Rodin 3D digital humans. He joined DeepSeek in 2023, where he served as a Senior AI Scientist and Project Lead for DeepSeek-VL. After leaving DeepSeek, he returned to Zhejiang University as a researcher and PhD supervisor while founding Simplex AI.
In response to questions about how Linger differs from existing bot frameworks, Zhang explained that the focus is on optimizing the model and harness together to enable local 125B model execution without per-token fees or data leaving the machine. While DS-Bot focuses on bot-to-bot and group collaboration, Linger centers on the interaction between the human user, their personal agent, and the local environment. The open-source nature of the project is rooted in a philosophy that intelligence should be distributed, ensuring that users retain full ownership of their context and data. This approach mirrors the historical shift from centralized computing to the personal computer, aiming to prevent AI from being strictly a rented token service hosted in data centers.
原文出处:量子位,https://www.qbitai.com/2026/11/511570.html
- 作者:数字Baby
- 链接:myaigc.top/article/41fb7c2d-0e98-811e-9aad-f22c0bc7ce2b
- 声明:本文采用 CC BY-NC-SA 4.0 许可协议,转载请注明出处。



