Simplex AI Open-Sources Personal Agent Linger and Local Inference Engine PolyStrata
Simplex AI founder Zhang Bo announced the open-sourcing of the personal agent Linger and the local inference engine PolyStrata, enabling agents and local models to run on personal computers. Linger's Another Me can monitor inboxes, calendars, and Slack and proactively move tasks forward. It also supports multi-agent collaboration, interaction with the local computing environment, and integration with tools such as Claude Code and Codex. The project reports a LoCoMo accuracy rate of 95%; Zhang Bo said PolyStrata can run the 125B-parameter Qwen3.8-Flash-Next with 64GB of RAM and on some configurations with 12GB of VRAM.
Why read this
SelectedAI score 78
The original article also presents Linger's proactive collaboration design, local deployment path, and benchmark data, offering a basis for comparing the trade-offs between cloud-based and on-device approaches for personal agents.
Topics
Body · AI translation
Former DeepSeek Researcher Open-Sources Local Version of Muse, Enabling Token Freedom — Zhidx Download Client Exclusive First Look — November 20, 2026, 17:42:40, from Beijing
Zhidx reported on November 20 that last night, former DeepSeek researcher and Simplex AI founder Zhang Bo (Tony Zhang) announced the open-sourcing of Linger, a personal agent that aims to free agents from dependence on cloud data centers and run directly on users’ own computers. The project quickly attracted attention after its release and has already surpassed 2,000 stars on GitHub.
▲Linger project’s GitHub homepage. Source: GitHub
Open-source address:
In a post on X, Zhang compared this shift to the birth of the personal computer. In the 1960s, computing power was concentrated in large computing centers. It was not until personal computers emerged that computing power gradually entered ordinary people’s lives.
Today, AI still runs primarily in data centers, and users must pay by the token to access it. Zhang hopes Linger can change this situation, allowing users to truly own their data, personal context, and the intelligence that powers AI.
▲Zhang Bo announces the open-sourcing of Linger. Source: X
The biggest highlight of the newly open-sourced Linger is its proactive assistant, Another Me. Unlike most agents, which require users to issue instructions, Another Me can continuously monitor a user’s inbox, calendar, and Slack, proactively alert the user when it finds matters requiring attention, and move subsequent tasks forward. At the end of each day, it also reviews the proactive messages it sent and adjusts its rules for future actions based on user feedback.
In addition to proactively handling tasks, Linger supports collaboration among multiple agents, can operate local files, browsers, and desktop applications, and integrates with tools such as Claude Code and Codex. Users can also choose local models, keeping the agent’s runtime environment and personal data on their own computers.
In the LoCoMo benchmark, which evaluates long-term conversational memory, Linger achieved 95% accuracy, outperforming several memory-enhanced AI systems, including Mem0 and MemOS.
▲Linger’s accuracy. Source: GitHub
To support running large models on personal computers, Simplex AI also simultaneously open-sourced the local inference engine PolyStrata. Zhang said that with 64GB of RAM, some computers equipped with 12GB of VRAM can run the 125B-parameter Qwen3.8-Flash-Next model; on an RTX 3060 with 12GB of VRAM, the model can achieve an inference speed of 45.4 tokens per second.
He also noted that, using the polystrata framework to place the MoE model’s hot experts in VRAM and its cold experts in system memory, speeds can reach 80 tokens per second, making it highly competitive.
▲Local inference engine PolyStrata. Source: GitHub
Online users showed varying degrees of curiosity. One user compared it with OpenClaw. Zhang responded that both are open-source personal assistants, but they have different focuses: OpenClaw places greater emphasis on integrating with commonly used chat applications and enabling collaboration through bots and groups, while Linger focuses more on collaboration between the user and the agent, as well as the agent’s ability to operate within the local computer environment.
▲Zhang Bo responds to a user’s comment about the differences between Linger and OpenClaw. Source: X
Another user asked about the differences between Linger and DeepSeek DS-Bot. Zhang said DS-Bot is a plugin developed on the basis of DeepSeek Harness, and its model can be accessed through an API configured by the user. Linger, by contrast, is not tied to a single harness: it can integrate with tools such as Claude Code, Codex, and Pi, and it also supports local models. Its focus is on enabling the model and harness to be optimized together.
▲Zhang Bo responds to a user’s comment about the differences between Linger and DS-Bot. Source: X
Zhang previously conducted AI research at several technology institutions. In 2013, he received his bachelor’s degree in Optoelectronic Information Engineering from Zhejiang University, and in 2019, he earned a PhD in Electronic and Computer Engineering from the Hong Kong University of Science and Technology. He subsequently joined the Visual Computing Group at Microsoft Research Asia as a senior researcher, where his research covered computer vision and image generation, and he participated in projects including CoCosNet, VQ-Diffusion, and Rodin.
In 2023, Zhang joined DeepSeek as an AI scientist and served as the project lead for the multimodal model DeepSeek-VL. After leaving DeepSeek, he joined Zhejiang University’s College of Computer Science and Technology as a Hundred Talents Program researcher and doctoral supervisor, and founded the AI company Simplex AI.
▲Zhang Bo’s work experience. Source: LinkedIn
This open-source release also represents a milestone achievement for Simplex AI in personal agents and local inference.
Conclusion: Agents Are Moving onto Users’ Local Devices
Currently, agents still largely depend on cloud-based models and computing power from data centers, with users accessing models through APIs. As on-device model inference capabilities continue to improve, running agents directly on users’ local devices is gradually becoming a new area of exploration. By open-sourcing Linger and PolyStrata, Zhang Bo is also attempting to optimize agents’ local performance and give users greater control over their own data and AI capabilities.
However, moving agents onto local devices is not easy. Given the constraints of device computing power and memory resources, running larger-scale models locally while maintaining agent responsiveness and task-execution capabilities remains a challenge that must be solved in this field.
Sources: X, LinkedIn, GitHub
“Special statement: The content above (including any videos, images, or audio contained therein) was uploaded and published by users of Dafeng Hao, a self-media platform operated by Phoenix New Media. This platform merely provides information-storage services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.”
AI Trends