OpenJev open-sources Jev-style semantic decisions on a single RTX 3090 with a frozen 4B model
alexcovo_eth · x · 2026-09-17
OpenJev is a new open-source project asking whether something like TypeSafe's closed Jev service—runtime-defined semantic decisions—can run at home on a single RTX 3090. It's available now with no waitlist, including a browser-based WebGPU demo.
The core insight: most agent decisions are small (routing, retries, evidence checks), and chat models waste time generating text that software immediately parses back into an if statement. OpenJev's baseline instead reads typed option probabilities directly from a frozen 4B model—no answer sentence, no JSON repair, no decoding loop.
Benchmarks use the same frozen 4B model, state, and 21 questions, measured separately and aligned at t=0 in replay. The project reproduces Jev's interface pattern, not its undisclosed model or training. It has already gathered 368 stars on GitHub.
More from coding & agent
- Unsloth Desktop Launches: Train & Run 500+ Local Models, 70% Less VRAM — danielhanchen · 2026-09-17
- Claude Code agent skills repo nears 1k stars with worktree workflow and before/after proof — Rasmic · 2026-09-17
- Economist Pushes Back on 'Read Your AI Code': Nobody Can Review 25,000 Lines — paulnovosad · 2026-09-17
- pen.dev launches an agentic design canvas that imports any website and connects to Claude, GPT, Qwen and more via MCP — yakuzeg · 2026-09-17
- Overprompted: An Overcooked-Style Co-op Game Where ChatGPT Plays via WebMCP — prd_008 · 2026-09-17
- Rotating Leaked GitHub Secrets: Deleting Commits Doesn't Scrub Git History — eyishazyer · 2026-09-17