Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableX: @GithubProjects→ original

Linger Runs 12 AI Models on Your PC: Up to 5x Faster and $0 in Token Costs

Cloud AI agents charge for every token, and the more files they read, the bigger the bill. Tony Zhang, one of DeepSeek's early developers, has open-sourced Linger, an agent that runs 12 models (including Qwen, DeepSeek, and Gemma) directly on your computer, so local work costs $0 in tokens. The up-to-5x speedup is the authors' own claim, and there are no independent measurements. If your API bill is hurting, test one of your processes locally.

AI-processed from X: @GithubProjects; edited by Hamidun News
Linger Runs 12 AI Models on Your PC: Up to 5x Faster and $0 in Token Costs
Source: X: @GithubProjects. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @GithubProjects ↗
Editor-in-Chief
Key metric #112 models run locally, including Qwen, DeepSeek, and Gemma
Key metric #2Data stays on the device only in local mode
Zero token fees do not mean zero costsyou need your own hardware and setup

Imagine you hand an AI agent a folder of contracts and six months of correspondence to go through. Every word it reads goes to the cloud, and the provider bills you for "tokens", the chunks of text you pay for. The more the agent reads and re-checks, the higher the bill.

The Linger team proposes the reverse model: an open personal agent that runs on your computer. It runs 12 leading models, including Qwen, DeepSeek, and Gemma, charges $0 in token fees for local inference, and, according to the authors, works up to 5 times faster. The project was announced by Tony Zhang, whom the announcement describes as one of DeepSeek's early key contributors and a former Microsoft Research Asia employee.

What Linger Does

According to the authors, Linger runs 12 models on supported hardware. Users are not tied to a single provider: you can pick a model for the task and for your graphics card or laptop.

Beyond running models, the agent takes over work that usually has to be done by hand:

  • It pulls context from the files and apps you have granted access to, so you don't need to copy documents into a chat;
  • It remembers preferences and past decisions, such as how you format reports or which stack you chose in previous projects;
  • It coordinates several agents to carry a task through to a result, rather than stopping at the first answer.

The authors describe it as a Grok-style experience, only on your own machine.

Why This Could Pay Off

The headline figures in the announcement are "up to 5 times faster" and $0 in tokens for local inference. The logic is simple. If an agent constantly re-reads files, summarizes long threads, and runs chains of dozens of calls, the cloud API bill grows quickly. A local model turns variable costs into fixed ones: you pay for the hardware once and stop counting tokens.

There is also an ideological argument. The authors recall that in the 1960s computing was centralized until the personal computer put power in everyone's hands. Today AI is trained in data centers and rented out by the token. Linger is trying to play the role of a "personal computer" for agents.

Limitations to Keep in Mind

The figures "up to 5 times faster" and "12 models" come from the authors' own announcement. The announcement does not say what the speed was compared against: a cloud API or another local wrapper such as Ollama. Results will depend heavily on hardware: large models on a weak laptop will run slowly, or may not run at all.

Zero token fees do not mean zero cost. You need video memory, electricity, and time for setup. Privacy is also conditional: the authors state plainly that data stays on the device only in local mode. If you connect external services or cloud models, information goes to them.

Security is a separate question. An agent with access to your files and apps presents a large attack surface. Before granting permissions, check what access it requests, how its memory is stored, and whether you can restrict access to individual folders.

What Developers and Product Managers Should Do

If your agent workflows are constrained by API costs and your data can't be sent outside, Linger is worth testing. Pick one repeating process, such as document review or work with a local repository, and compare quality, speed, and total cost against the cloud version. Check the license and the list of supported hardware in the repository. If "up to 5 times faster" holds up even on your configuration, local agents gain another strong argument.

Story

Key Questions Answered

What Linger Does?

According to the authors, Linger runs 12 models on supported hardware. Users are not tied to a single provider: you can pick a model for the task and for your graphics card or laptop. Beyond running models, the agent takes over work that usually has to be done by hand: It pulls context from the files and apps you have granted access to, so you don't need to copy documents into a chat; It remembers preferences and past decisions, such as how you format reports or which stack you chose in previous…

Why This Could Pay Off?

The headline figures in the announcement are "up to 5 times faster" and $0 in tokens for local inference. The logic is simple. If an agent constantly re-reads files, summarizes long threads, and runs chains of dozens of calls, the cloud API bill grows quickly. A local model turns variable costs into fixed ones: you pay for the hardware once and stop counting tokens. There is also an ideological argument.

What are the key details on Limitations to Keep in Mind?

The figures "up to 5 times faster" and "12 models" come from the authors' own announcement. The announcement does not say what the speed was compared against: a cloud API or another local wrapper such as Ollama. Results will depend heavily on hardware: large models on a weak laptop will run slowly, or may not run at all. Zero token fees do not mean zero cost. You need video memory, electricity, and time for setup.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…