APUS 开源国内首批Jev跨平台复现：国产模型实现秒级决策
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - APUS open-sourced fast-browser-use, a cross-platform Agent Skill that reproduces Jev-style rapid decision-making with local models. It enables offline browser agents to select actions through single-pass scoring instead of autoregressive text generation, reducing latency, hallucination risk, API costs, and data exposure.
- Uses a local Qwen3.5-9B model to score numbered browser actions directly from single-token logits, avoiding generated selectors and formatting errors.
- Implements KV-cache broadcasting and concurrent batch evaluation for high-frequency classification, selection, scoring, and verification tasks.
- Supports macOS, Linux, and Windows with or without a GPU, and is released under the MIT license.
- APUS reports median completion of about 18 seconds for a Wikipedia retrieval task on an Apple M2 Pro, while form-filling and site navigation took roughly 3 seconds, fully offline.
Sources (1)
APUS 开源国内首批Jev跨平台复现：国产模型实现秒级决策
TL;DR - APUS open-sourced fast-browser-use, a cross-platform Agent Skill that reproduces Jev-style rapid decision-making with local models. It enables offline browser agents to select actions through single-pass scoring instead of autoregressive text generation, reducing latency, hallucination risk, API costs, and data exposure.
- Uses a local Qwen3.5-9B model to score numbered browser actions directly from single-token logits, avoiding generated selectors and formatting errors.
- Implements KV-cache broadcasting and concurrent batch evaluation for high-frequency classification, selection, scoring, and verification tasks.
- Supports macOS, Linux, and Windows with or without a GPU, and is released under the MIT license.
- APUS reports median completion of about 18 seconds for a Wikipedia retrieval task on an Apple M2 Pro, while form-filling and site navigation took roughly 3 seconds, fully offline.