APUS Open-Sources Early Jev Replication for Offline Browser Automation

Jev, named after 19th-century economist William Stanley Jevons, alludes to the Jevons paradox: a sharp fall in the cost of intelligence use will bring a surge in intelligence consumption. The model targets a high-frequency need in agent operations, namely classification, selection, scoring and truth judgment, returning typed answers and confidence probabilities rather than generating text token by token. The industry view is that this route validates a fast-slow division of labor in agent architecture: expensive large models handle slow thinking such as planning, while high-frequency, atomic fast judgments are left to lightweight decision models. The engineering layer that assembles and schedules this division is called Harness.

APUS says it fully implemented Jev's underlying decision paradigm and applied it to the most representative browser automation scenario. Based on the input paradigm and evaluation logic shown in Jev's public documentation, APUS analyzed its core logic of skipping autoregressive decoding and directly scoring from hidden states. In its open-source project, APUS replicated single-token logits fast decision-making, skipped autoregressive decoding, and implemented Jev's KV-Cache broadcast and concurrent batch evaluation mechanisms. In browser automation, fast-browser-use organizes real visible and interactive elements on a page into a numbered candidate action set. A locally running Qwen3.5-9B model then completes the where-to-click, what-to-choose decision through a single forward computation, mechanically eliminating the possibility of the model generating incorrect selectors or format hallucinations.

APUS's measured data shows that on an Apple M2 Pro consumer laptop, an agent relying on local model inference completed a real Wikipedia retrieval task entirely offline with a median time of about 18 seconds. Form filling and in-site navigation tasks took about 3 seconds. The model scored only four times per task, with zero cloud calls, zero API costs and user data kept on the machine.

APUS AI Lab said the open-source replication has three meanings: first, it verifies the engineering feasibility of using a model as a Harness; second, it proves that beyond closed-source APIs, open weights plus edge computing can also support this paradigm, with data remaining local to meet data security requirements in government, enterprise and financial scenarios; third, it provides the industry with an independent verification reference with transparent data and reviewable code. Previously, Jev's performance data all came from TypeSafe's own official self-testing.

According to the information, fast-browser-use has been packaged as a standard Agent Skill, and developers can integrate it into mainstream coding agents such as Claude Code, Codex and OpenCode with a single command. APUS AI Lab also said it will continue to open-source new Skills around desktop GUI operations, local operating system interaction and other scenarios.

Public information shows APUS was founded in 2014. Its AI products and services cover 2.4 billion users in more than 200 countries and regions. In July 2026 it released a new brand positioning as a global AI Agent ecosystem builder. Its core technology route is to pursue the most advanced LLM technology, go beyond pure model competition, deepen implementation applications and focus on real industry value output. In the future, APUS will continue to use its self-developed large models as a foundation, focus on AI Agent core technology innovation, and build an Agent product matrix covering financial risk control, intelligence analysis, content creation, healthcare and other scenarios.

The project address is https://github.com/APUS-AI-Lab/fast-browser-use. The announcement was reported by QbitAI, which said the article was provided by APUS and reposted with authorization, with views belonging to the original author.