APUS open-sources a cross-platform Jev reproduction that runs browser agents fully offline

On September 19 the AI lab of Chinese company APUS published what it calls one of the world's earliest independent open-source reproductions of Jev, packaging the work as a ready-to-use Agent Skill named fast-browser-use that runs entirely on local models with no cloud dependency. QbitAI reported the release on September 20.

The project is deliberately broad in reach: it supports macOS, Linux and Windows, runs on GPU servers as well as Macs and PCs without a GPU, and ships all of its code under the MIT licence so developers can reuse it freely.

Jev is named after the nineteenth-century economist William Stanley Jevons, a nod to the Jevons paradox — as the cost of using intelligence collapses, the volume of intelligence consumed rises. The model targets the most frequent requests inside an agent loop, namely classification, selection, scoring and true-or-false judgement, returning typed answers with confidence probabilities instead of generating text token by token. Industry observers read the approach as a fast-and-slow division of labour, where expensive large models handle planning while high-frequency, atomic decisions go to a lightweight decision model; the engineering layer that assembles and schedules the two is known as the harness.

APUS says it worked from Jev's public documentation, analysing the input format and evaluation logic to identify a core trick — skipping autoregressive decoding and scoring directly on hidden states — and reproduced single-token logits decision-making together with KV-cache broadcasting and concurrent batched evaluation.

In the browser automation setting, fast-browser-use turns every visible, interactive element on a page into a numbered set of candidate actions, then has a locally running Qwen3.5-9B model pick where to click or what to choose in a single forward pass. Because the model selects from a prepared list rather than writing a selector, generating a malformed selector or hallucinating a format is structurally impossible.

The published measurements are the most persuasive part of the story. On a consumer Apple M2 Pro laptop, the agent completed a real Wikipedia retrieval task fully offline with a median time of about 18 seconds and only four model scoring calls per task. Form filling and in-site navigation took roughly three seconds each, with no cloud calls, no API fees and no user data leaving the machine.

APUS frames the release as proving three things — that the model-as-harness route is engineering-feasible, that open weights plus on-device compute can support the paradigm while keeping data local for government, enterprise and financial users, and that the industry now has an independently reviewable reference, since Jev's performance figures had previously come only from TypeSafe's own testing.

Adoption should be straightforward: fast-browser-use is packaged as a standard Agent Skill that installs with one command into mainstream coding agents including Claude Code, Codex and OpenCode, and APUS says further skills covering desktop GUI operation and local OS interaction will follow. Public information describes APUS as founded in 2014, with AI products and services reaching 2.4 billion users across more than 200 countries and regions.

What to watch next is whether this independent reproduction can be benchmarked head to head with Jev's official numbers on more real tasks, and whether on-device decision models can keep second-level latency and low scoring counts as they move into desktop GUI and file-operation scenarios.