AI Products量子位

3x Faster Response Than Claude Code, Truly Infinite Context: APUS AI Open-Sources Reflex Agent

量子位Oct 14, 20261 min read

AI Summary

APUS AI (Qilin Hesheng Network Technology Co., Ltd.) has open-sourced its personal agent product, Reflex Agent. Built on its proprietary open-source decision model Reflex and the Continuum memory engine, the product reportedly responds up to three times faster than mainstream agents such as Claude Code and Codex in the company's tests.

AI Insights

APUS AI's open-sourcing of Reflex Agent under the MIT license signals a push toward decision-model-driven agent harnesses, shifting control logic from hand-written prompts and code to a dedicated model. The claimed 3x speedup and 'infinite context' rely on self-reported internal tests and a memory design that reduces reliance on compressed summaries rather than expanding the model's actual input window, so independent benchmarks are needed before drawing conclusions. Overall, this reflects a broader trend in which Chinese AI firms compete on agent architecture and open ecosystems, not just base model scale.

Analysis written by AI from the original article. Check it against the source.

Translated text

APUS AI has released Reflex Agent under the MIT license, offered as a desktop app, web application, and SDK so developers can try, reproduce, and extend it. In this product, the decision model Reflex handles decision work such as model routing, skill selection, result review, and execution supervision. The company calls this operating mechanism 'Loop Working,' describing a cycle of 'decide, execute, feedback, re-decide' that keeps tasks moving forward. On memory, Reflex Agent uses the Continuum memory engine to accumulate task information without limit, preserve original content and its associations losslessly, and let relevant historical information participate in decisions in its original form wherever possible. This allows the agent to retrieve and call historical information on demand during long-running tasks, reducing reliance on compressed summaries. The 'infinite context' referred to here does not break the input length limit of a single model call; rather, the agent's memory and task duration are no longer fully bound by a single context window. The release is described as an important step in APUS AI's exploration of Loop Working and personal agents. The project is available at github.com/APUS-AI-Lab/APUS-Reflex-Agent, and the decision model at huggingface.co/apus-ailab/APUS-Reflex.

Related News

AI Products36氪

Geeks Build Cabbage-Priced Biped Robot for Under 3,000 Yuan, with Old Phones as Key Contributors

UCLA's RoMeLa lab has developed PhoneBot, a 1.8 kg biped robot with 13 degrees of freedom and a materials cost of about $400 (roughly 2,677 yuan). The robot is powered by a smartphone that serves as its sensor suite, onboard computer, and interface, and by using used older phones, total cost stays under 3,000 yuan.

AI Products36氪

Alibaba, ByteDance and Meta Ignite Personal Agent Race: No Table Without a Good Model

In late September, Alibaba announced at its Yunqi Conference that Qwen (Tongyi Qianwen) would accelerate personal agent development, and on October 13 it unveiled 'Qwen Go!', a personal assistant in limited invitation testing. ByteDance (Doubao), Tencent and others have also launched personal agent projects, sparked by Meta's Muse. On the model side, Meta's Muse Wonder and Alibaba's QwenGyre reinforcement-learning framework aim to strengthen long-running agent capabilities.

AI Products36氪

Tech Giants at Home and Abroad Scramble for the Muse Entrance, But the Personal Agent Remains a False Proposition

A month after Meta launched the personal agent Muse, touted as the most captivating product since ChatGPT, ByteDance, Tencent, Alibaba, and Ant Group have rushed out similar offerings. However, Muse's daily net new users fell 62.4% from the September peak, and analysis suggests the model struggles to fit the domestic Chinese market in both product logic and business model.