APUS reverse-engineers Jev into a fully offline browser agent
Jev's whole pitch was that decisions do not need generated text. Someone has now rebuilt that pitch with open weights, on a laptop, with the benchmark table to go with it.
China's APUS AI Lab published fast-browser-use, an open-source reproduction of Jev's decision mechanics that runs entirely on local Qwen3.5 weights — no cloud endpoint, no API bill, MIT licensed. The project is not a wrapper around TypeSafe's hosted model, which is what Browser Use's own Jev Ultrafast was. APUS says it worked from Jev's public documentation and evaluation logic to infer the core trick — skip autoregressive decoding and score the answer directly off the hidden state — then rebuilt it: candidate actions on the page are mapped to single vocabulary tokens, and one forward pass returns the softmax probabilities over the legal choices.
The engineering follows from that. A scanner reads the rendered DOM and extracts only visible, interactable elements into numbered candidate tuples — click this button, select this option, type into this field, or declare the task done. The model never writes a selector or a Playwright call, so an invented CSS path is structurally impossible rather than merely unlikely. Structural actions use discrete logits scoring; generative text is invoked only when the chosen action is to type something. Page context is prefilled once and the KV cache is broadcast across candidates, so scoring many fields or options collapses into one batched pass. Jointly scoring the candidates and the completion signal cut inference passes per task from 14 to 4.
The published numbers are consumer-hardware numbers, which is the point. On an Apple M2 Pro with 32 GB of unified memory and 4-bit MLX weights, a live Wikipedia task — start at the main page, find and open the article on Python, verify the final URL and title — had a median of 30.09 seconds on Qwen3.5-9B and 18.90 seconds on Qwen3.5-35B-A3B, in four scoring steps. A workspace settings form (name, timezone dropdown, digest toggle, save) finished in 12.00 seconds, a local reading-room navigation in 5.43 seconds, and two short hops in 8.10 and 7.31 seconds. Every run is checked against an external assertion on URL or title rather than trusting the model's own "done" — a convention APUS states outright, since the completion signal is a hypothesis, not evidence. The 9B checkpoint is about 5.95 GB in 4-bit; unquantized 9B wants roughly 18 GB, and the 35B-A3B wants an 80 GB card.
Why it matters beyond one repo: Jev ships as a closed API with no weights and no technical paper, so every performance claim about the architecture has come from TypeSafe itself or from projects pointing at its endpoint. This is the first independent rebuild that anyone can download and re-run — and the first to put the paradigm in a form that keeps page contents on the machine, which is the version enterprise and financial buyers actually need. It also lands in a fast-filling genre: Cua published a 706,048-parameter form-filling model two days ago — Cua's 2.8 MB open model fills forms without an LLM call — and an open Laya checkpoint shipped with head-to-head scores against Jev. We covered the credit fight that runs underneath all of it in The Jev idea was published a year ago — open, and unnoticed.
The caveats are the vendor's own. The latency table is MLX on Apple Silicon; the PyTorch path on Linux recomputes the prompt for each score and does not reuse a prefix between decisions, and it does not yet load 4-bit weights or shard across GPUs. The benchmarks were run by the team that wrote the harness, and the scoring harness is the part most likely to flatter a model. Nothing here has been reproduced by a third party.
What to watch: whether an independent lab re-runs the Wikipedia task, and whether the follow-up skills APUS promised — desktop GUI control, local OS interaction — keep the same local-only constraint.
If a laptop can drive a browser with no cloud call at all, what is left for the paid agent APIs to sell? Tell us in the comments.
Sources: APUS-AI-Lab/fast-browser-use (GitHub) · fast-browser-use performance measurements (GitHub)
