User repository on GitHub
Same as jev-ultrafast but using Laya
- Forks
- 43
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
Laya Ultrafast
A clone of browser-use's jev-ultrafast that swaps the hosted Jev decision call for Laya running locally through MLX, keeping the browser agent, action space and executor but making no paid decision API calls. Apple Silicon only.
- Stars
- 253
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- ipenywis
- Last verified
- Oct 2026
About this project
A local, open-weight port of browser-use/jev-ultrafast. jev-ultrafast asks TypeSafe's Jev, a hosted model, to choose each browser action. This port replaces that API call with Laya, an open-weight typed-decision model that runs on your Mac: No decision API and no per-step cost. Each decision is a local forward pass. The median is about 33 ms on an M1 Max.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Local decisions are a forward pass with a median around 33 ms on an M1 Max and no per-step cost, against the hosted call it replaces.
- One text-model call per run turns the goal into field values, the item to open and a finish condition; point it at Ollama and the agent works offline.
- Laya is asked narrow questions only, then rules fill values, pick from options that appeared and submit; targets are still observed elements.
- DECISION_MODEL=typesafe restores the original hosted Jev policy unchanged, so both decision paths live in the same codebase.
Watch out
This is a port, not a new agent: the browser agent, DOM extraction, executor, safety checks, inspector and performance work are upstream's, and the README sends you to browser-use/jev-ultrafast for how any of it works. It is unusually clear that the Laya policy is new and tested on few sites, and that the published measurements are a small set of repeated tasks rather than a reliability benchmark: the flight run passed five out of five, the local fixtures two out of two each, and Skyscanner was blocked by its robot check, which the agent deliberately does not try to bypass. A DONE decision is explicitly not proof of success, so the examples verify the final page independently. Planning still depends on the text model, and the default sometimes returns malformed JSON, so the planner retries up to three times. The default checkpoint has a 1,024-token context and the English Laya variant only 512. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Because it is a per-step API call, and this path makes no decision API calls at all — Laya runs as a local forward pass with a median around 33 ms on an M1 Max. The cost moves to your hardware and to different behaviour: since Laya does not reliably answer the open question of what the browser should do next, the policy is narrow Laya questions plus three site-agnostic rules. That is a real change in how decisions get made, not just where they run.
No. Laya runs through MLX, which needs an M-series Mac, macOS 14 or later, and Python 3.12 or newer for this project. The README points at the upstream Laya repository for other platforms. Chrome also connects through Browser Harness with remote debugging enabled at chrome://inspect, and there is a --doctor command to run when that connection fails.
Almost. Decisions are local, and if you point the text-model base URL at a local Ollama server then no key is needed and the only network traffic left is the sites being browsed. You do have to download the checkpoint once with the hf command; after that it loads from the Hugging Face cache and needs no network for decisions. Out of the box the text model defaults to OpenRouter, so it is not offline until you change that.
Upstream's limits carry over unchanged: shadow roots, frames, canvas, uploads, pop-up tabs, nested scrolling and arbitrary keyboard widgets are out of scope. Add the risk of a bot check — the Skyscanner example hits one in headless testing and the agent refuses to solve it, so run it in your everyday Chrome and handle the check yourself. Also note the flight examples never select or book anything, and the policy has been exercised on few sites.
Set DECISION_MODEL=typesafe and provide TYPESAFE_API_KEY and TYPESAFE_MODEL. Both decision paths live in the same codebase, so the original jev-ultrafast policy runs unchanged and no other edits are needed.
It installs laya-mlx as a dependency and uses the aac6fef/laya-typed-decisions-mlx checkpoint, 421M parameters with a 1,024-token context. You fetch it once with the hf command, after which decisions load from the Hugging Face cache with no network.
Chrome connects through Browser Harness, exactly as in jev-ultrafast. You enable remote debugging at chrome://inspect/#remote-debugging and run browser-harness --doctor if the connection fails.
More Jev resources like this
1.7kGitHub stars
Open-weight 4B and 9B models from TokenRhythm: NeoHorse-1 for tool use, coding, and instruction following, plus NeoHorse-Jev-4B for prefill-only Choice, Noul, and Score decisions.
Jev Models & Reimplementations#community#jevbest
Community
TokenRhythm/NeoHorsePython
TokenRhythm/NeoHorsePython1.5kGitHub stars
A 0.8B decision model that answers first and escalates only when unsure, plus nine LoRA adapters of roughly 41 MB each for guard, triage, tool choice, spam, and similar classification work.
Jev Models & Reimplementations#community#jevbest
Community
firelex/jeffPython
firelex/jeffPython2kGitHub stars
DeepOpen is a non-autoregressive decision engine built on Laya that classifies into developer-defined types in one forward pass, with a language router picking between English, multilingual, and typed-decision checkpoints.
Jev Models & Reimplementations#community#jevbest
Community
deepopen-com/deepopenPython
deepopen-com/deepopenPython715GitHub stars
Decision models that answer several choice, score, and noul questions over a state with probabilities in one forward pass, with training, two inference backends, calibration tooling, and the benchmark suite in the repo.
Jev Models & Reimplementations#community#jevbest
Community
InternLM/Intern-DecisionPython
InternLM/Intern-DecisionPython402GitHub stars
Decision models that return typed probabilities for candidate labels from a state, instructions, and options, shipped as LoRA adapters plus a trained decision head over pinned Qwen base weights.
Jev Models & Reimplementations#community#jevbest
Community
Zefan-Cai/Open-JevPython
Zefan-Cai/Open-JevPython363GitHub stars
Open-weight 2B, 4B and 9B decision models that extend Jev's typed question format to photos, so an image plus a record can be checked against options you supply, with a probability on each and an explicit can't-tell. Runs on your own hardware.
Jev Models & Reimplementations#community#jevbest
Community
mohit67890/imajevPython
mohit67890/imajevPythonFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.