BeijingA 4.5B-parameter model decides each agent step, not a giant one.
Long tasks overflow context windows, and the usual fix, compressed summaries, loses detail, APUS says.
A Qwen3.5-based decision model picks each action, and APUS claims 2-3x faster page responses in its own tests.
Developers can test the MIT-licensed code now, but nobody has yet verified it on complex business tasks.
How each outlet framed it
- Zhidx leans favorable
- emphasizes APUS's pivot from consumer apps to LLMs and claims 2-3x faster page response than rival agents via a 4.5B decision model
Sources: Zhidx