Skip to content
DeepSeek Alumni Launches Linger: Open-Source Personal Agent Runs 125B Model on 12GB GPU

DeepSeek Alumni Launches Linger: Open-Source Personal Agent Runs 125B Model on 12GB GPU

Former DeepSeek-VL lead Tony Zhang unveils Linger, an open-source personal AI agent that runs a 125B-parameter MoE model at 45.5 tokens/sec on a single RTX 3060 (12GB VRAM) via the PolyStrata inference engine, featuring local-first memory, proactive scheduling, and implicit reinforcement learning.

Machine Learning Algorithms & Natural Language Processing
Machine Learning Algorithms & Natural Language Processing
Machine Learning Algorithms & Natural Language Processing

Former DeepSeek V1 author and DeepSeek-VL project lead Tony Zhang (张博) has announced Linger, an open-source Personal Agent designed for local deployment on consumer hardware. The system centers on an agent called "Another Me" that lives on the user's PC and integrates with existing browser sessions, email, calendar, and chat logs without requiring new credentials.

Core Technical Innovation: PolyStrata MoE Inference Engine

Linger's key enabling technology is PolyStrata, a Mixture-of-Experts inference engine that implements intelligent tiered storage across GPU VRAM, system RAM, and SSD. For a large MoE model, the most frequently activated experts reside in GPU memory, less active experts in RAM, and rarely used experts on SSD. Data moves dynamically between tiers when experts are needed, avoiding the requirement to fit the entire model in VRAM.

Benchmark results cited in the article:

125B-parameter Qwen3.8-Flash-Next on RTX 3060 (12GB VRAM) + 64GB RAM: 45.5 tokens/sec

Same model on RTX 5090 (32GB VRAM): 179 tokens/sec

Model Autopilot

Linger includes a "Model Autopilot" that auto-detects hardware (GPU, VRAM, RAM) and selects from 49 language models, two inference engines, multiple quantization schemes, and context lengths to install the highest-performing combination. SimplexAI's scoring shows:

8GB GPU + 32GB RAM → score 35.4 ≈ Gemini 2.5 Pro

Mac 24GB unified memory → score 51.9 ≈ Claude Opus 4.5

12GB GPU + 32GB RAM → score 57.3 ≈ Claude Opus 4.6

Agent Capabilities

Memory Hub (Wiki): Each night, Linger ingests inbox, calendar, chat logs, and meetings into a local knowledge base. On the LoCoMo long-conversation memory benchmark, it achieves 95% accuracy, surpassing Mem0 and MemOS.

Proactive Scheduling: The agent observes information flow and decides when to interrupt based on urgency/importance. Example: a client email requesting a quote by Friday triggers a reminder on Wednesday morning; a meeting time change triggers immediate notification.

Implicit Reinforcement Learning: The system learns from user feedback without explicit programming. If the user replies "I got that invoice last week, should have told me then," the agent logs the signal. After a second similar occurrence, it auto-generates a rule: "Supplier invoice → immediate alert." Conversely, two ignored delivery notifications teach it to stay silent. Users can also give direct one-shot instructions like "Notify me instantly on client replies" or "After 9 PM, don't alert on meeting changes."

Desktop/Browser Control: Linger drives already-logged-in Chrome and local apps, no re-authentication needed.

Agent Teams (Super Agent): Complex tasks spawn specialized sub-agents that collaborate while the user stays in a single session.

Voice Interaction: Real-time meeting transcription; post-meeting Q&A and action-item follow-up.

Philosophy & Differentiation

1960s computing was centralized until PCs put power in everyone's hands. Today AI trains in data centers and rents tokens back to you. We believe intelligence should be distributed. Your context, your data, and the intelligence driving them should be fully yours.

When asked how Linger differs from DeepSeek's DS-Bot, Zhang replied: Linger co-optimizes model and harness so a 12GB card runs 125B locally with zero token cost and no data leaving the machine. Product-wise, DS-Bot focuses on bot-to-bot and bot-to-group collaboration; Linger focuses on human-to-self-agent and agent-to-local-environment interaction.

Team Background

Tony Zhang: B.S. Zhejiang University, Ph.D. HKUST (2019), Microsoft Research Asia (VQ-Diffusion, Rodin 3D), DeepSeek (DeepSeek LLM V1 series, DeepSeek-VL project lead), now Zhejiang University "Hundred Talents" researcher and Simplex AI founder. The PolyStrata engine was built by an 18-year-old team member.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Mixture of Expertsopen-source AIOn-Device AIPersonal AI AgentLocal LLM InferenceImplicit Reinforcement LearningMemory Augmented AgentsPolyStrata
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.