Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents (Nov 23, 18:30 MSK): 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseCode / SDK · ActionableX: @gkxspace→ original

DeepSeek Veteran Open-Sources Linger: A Local AI Agent That Responds Up to 5x Faster

Linger is an open-source AI agent that runs on your own computer, not in the cloud with per-token billing. The team behind it is led by one of DeepSeek's earliest key contributors, according to the review's author. You can tell the main assistant "bring in a frontend developer and an independent QA" and the result arrives only after review and fixes. The PolyStrata engine runs 12 open models locally and, in special configurations, speeds up responses by up to 5x. Check that figure on your own hardware.

AI-processed from X: @gkxspace; edited by Hamidun News
DeepSeek Veteran Open-Sources Linger: A Local AI Agent That Responds Up to 5x Faster
Source: X: @gkxspace. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @gkxspace ↗
Editor-in-Chief
Key metric #1Roles and handoff order are defined in a team.md file
Key metric #2Long tasks go to a background Task, leaving the chat free
Key metric #3Codex and Claude subscriptions or local Qwen, DeepSeek, and Gemma models can be connected

Linger is an open-source AI agent that lives on your own computer, not in someone else's data center. Usually it works the other way around: the model runs on servers, you pay for every token (a chunk of text the model reads or writes), and your files go to the cloud. The Linger project from SimplexAI Labs offers a different setup: a multi-agent workspace that runs locally.

According to the review's author on X, the project is led by one of DeepSeek's earliest key contributors. Alongside the agent, the team has also released PolyStrata, an engine that runs models on your machine. The "up to 5x" in the headline refers specifically to the response speed of local models in specially chosen configurations, not to an overall quality gain.

A note about the source. This is not an official press release but an enthusiastic review from a user who, as he writes, tested Linger for several days. We did not independently verify the project leader's background. The authors attribute the speed figures to "special configurations," which means they represent a maximum, not an average.

The developers explain the idea through history: in the 1960s computing was centralized until the personal computer gave power to everyone. Today AI is trained in data centers and "rented out" by the token. Linger is an attempt to repeat that shift with the PC.

A team of agents instead of a single chat

The most practical part is multiple agents in one "room." You tell the main assistant in plain language: add a frontend engineer and an independent QA, draft a team.md file, and give me the result only after QA finds the weak spots and the engineer fixes them.

The team.md file works as the rulebook of a small team: it records who is responsible for what and who receives the work next. This resembles ordinary code review. The author does not check their own work; a second participant specifically looks for what broke. For a developer the scheme is familiar, except the contributors are models. The point is that the reviewing agent does not simply agree with the first answer out of politeness.

Background tasks and model choice

The second scenario is background "doubles." Long jobs, such as a shot-by-shot script for a video or analysis of a large set of materials, can be sent to a separate Task. While it runs, you can keep asking in the main window which stage it has reached and whom to nudge next. When the task finishes, the finished document appears in the chat with a preview. Long operations do not block the conversation.

The third point is flexibility with models. Linger can connect Codex and Claude subscriptions, meaning you can use already-paid cloud models as the "brain" of the agents. It also works with local model execution, handled by PolyStrata. The claimed support covers 12 open models, including Qwen, DeepSeek, and Gemma. In the best configurations, responses come up to 5x faster. The post does not show which specific models, what hardware, or which modes stand behind that number. You will have to compare against llama.cpp and other engines on your own laptop.

What to do in practice

For developers and product managers, Linger is interesting for three reasons. First, it is a ready-made workspace rather than a set of libraries: multi-role work, background tasks, and result previews are already assembled. Second, the hybrid scheme lets you keep sensitive tasks on a local model and send complex ones to cloud subscriptions. Third, the open code can be read, verified, and adapted to your own team.

There are limitations too. Local models on an ordinary laptop still trail flagship cloud models in quality. Several agents in one room consume more resources than a single one. The claimed "up to 5x" needs to be reproduced on your hardware before you build plans around it. The project is new, so it is too early to judge stability, the security of agents' file access, or the quality of documentation.

If you have long wanted to assemble a personal agent team without renting someone else's servers, run Linger on a test project. Start with a simple pair of "executor and independent QA" and see how much the results actually improve. Move real work over only once you see the effect and have measured the speed on your own computer.

Story

Key Questions Answered

What are the key details on A team of agents instead of a single chat?

The most practical part is multiple agents in one "room." You tell the main assistant in plain language: add a frontend engineer and an independent QA, draft a team.md file, and give me the result only after QA finds the weak spots and the engineer fixes them. The team.md file works as the rulebook of a small team: it records who is responsible for what and who receives the work next. This resembles ordinary code review.

What are the key details on Background tasks and model choice?

The second scenario is background "doubles." Long jobs, such as a shot-by-shot script for a video or analysis of a large set of materials, can be sent to a separate Task. While it runs, you can keep asking in the main window which stage it has reached and whom to nudge next. When the task finishes, the finished document appears in the chat with a preview. Long operations do not block the conversation. The third point is flexibility with models.

What to do in practice?

For developers and product managers, Linger is interesting for three reasons. First, it is a ready-made workspace rather than a set of libraries: multi-role work, background tasks, and result previews are already assembled. Second, the hybrid scheme lets you keep sensitive tasks on a local model and send complex ones to cloud subscriptions. Third, the open code can be read, verified, and adapted to your own team. There are limitations too.

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.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…