openjev

Measured replay: typed decisions appear together while JSON streams token by token — Image from github.com

TheoLeeCJ/SemIf-OpenJev on GitHub

Stars: 4,142
Forks: 277
Language: Python
License: MIT
Last push: 23 Sep 2026

Read from GitHub on 24 Sep 2026. Counts change daily.

Reported by the author

JSON vs direct: 5.21x slowercompleting the array took 5.21× as long as direct readoutgithub.com
Peak throughput: 20.03 dec/sParallel suffixes | **20.03** | **38.8 s**github.com

SemIf, formerly called OpenJev, is an independent attempt to reproduce Jev’s interface on hardware anyone can own: a single RTX 3090 running a frozen Qwen3.5-4B model. It is not affiliated with TypeSafe and does not reproduce Jev’s own model or training; it copies the input and output shape, unstructured state plus runtime-defined criteria and typed options in, a probability per option out.

The project reads option probabilities directly from the model in one forward pass, skipping the answer sentence and JSON parsing a chat model would need. Benchmarked against the same model generating a compact JSON array of yes and no answers, direct readout took 1.023 seconds against 5.332 seconds for the generated array on 21 binary criteria, a gap the author frames as a systems comparison rather than a claim that the two readouts are semantically equivalent: the two methods agreed on 18 of 21 decisions. On a 37-state, 21-criterion workload, reusing a prefilled state pushed throughput from 2.33 decisions per second with fresh scoring to 20.03 decisions per second with parallel suffixes.

The repo ships an MLX backend for Apple Silicon, a llama.cpp path for CPU-only runs, and a browser demo that needs no waitlist. Recent commits added PyTorch and MPS scoring and per-workload temperature calibration.

Open on GitHub

Jev Ultrafast (Browser Use)

Browser Use's fast browser agent where Jev picks the operation and the DOM element in a single request; a small LLM is only invoked to write text when typing is required. Demoed on a Google Flights search.

browser-useOpen

19k starsFlight search 7.1 s

ProjectTyped tool dispatch+2

jev-trader

Makes one AI trade decision every Monad block on the Kuru MON-USDC market using Jev. The public deployment runs in dry-run mode with a mock model.

jarrodwattsOpen

2.2k starsLoop, mock model p50 100 msDecision window ~300 ms/block

ProjectConfidence-gated actions+1

jevlike

Open reimplementation that trains a small model to choose among a changing list of text options, emitting one probability per option in a single forward pass. Ships Doom, chess and Wikispeedia demos.

vinnylarouge1.3k starsOpen

Project

Ran Jev against an existing classifier eval that previously used Gemini 2.5 Flash Lite

Vercel CTO reports Jev saturated an existing classifier eval that had used Gemini 2.5 Flash Lite and ran about 6x faster. Original post; TypeSafe's quote-tweet is a separate item.

Malte Ubl (@cramforce)Open

220k viewsSpeed vs Gemini 6x faster

ProjectHierarchical classification+1

Agent Desktop

Rust desktop automation that reads apps through the OS accessibility tree instead of pixels. Its jev-desktop skill takes a plain-language goal and has Jev pick one operation and one target element per turn, so the 150-element tree never enters the agent's context.

lahfirOpen

1.6k starsToken reduction 78-96%

ProjectTyped tool dispatch+1

typesafe-computer-use

macOS computer-use loop that OCRs the screen, classifies the next action with Jev, then clicks. Author reports roughly $0.0002 per step.

Aaron Levin (awlevin)Open

935 starsCost/decision $0.0002, 155xLatency 0.13-0.38s

ProjectReal-time control+1

Menu