Open Jev reproductions ship in a day, typed decisions on one GPU
A day after TypeSafe's hosted Jev, open self-hostable versions arrived: in-browser on WebGPU, on a single RTX 3090, under Apache licenses.
Copy markdown
The demo runs both modes in your browser
OpenJev.com loads a pinned GGUF (Qwen3 0.6B, MiniCPM5 2B, or Qwen3.5 4B) over WebGPU and lets you A/B the two paths on one model: read typed option logits directly, or generate the same choice as JSON token-by-token. No backend, ~1.56GB cached in-browser. It hit 500+ points on Hacker News in a day.
One forward pass instead of a JSON round-trip
The shared trick: restrict the next-token distribution to your allowed options and read the probabilities once, skipping decoding entirely. ikermoel's Apache-2.0 lib clocks 92.9% on RACE-H at 2.5x fewer tokens and 2.3x faster than batching on Qwen3.6-27B, and 84.2% on MMLU. Typed, calibrated answers from any open-weights model, no hosted API.
It runs on the GPU you already own
TheoLeeCJ/openjev reads logits from a 4B model on a single RTX 3090 and reuses one long context across 21 criteria at once (~900 stars in two days). Verdict-open-jev goes fully non-autoregressive on a 151M ModernBERT with RLCD-calibrated uncertainty and a WebGPU playground. Support-ticket routing and bulk classification, local.
Bespoke ships an open recipe, not just weights
Bespoke Labs' Nimble drops open data, an open model, and the full open recipe for a Jev-style decider, reproducible end to end. The honest catch across every repo: none claims to match TypeSafe's hosted Jev on quality, so temperature-scale and benchmark on your own data before you trust the numbers. A curated 'awesome-jev' list is already tracking who built what.
Update: Codex CLI 0.155.1 unbreaks local TUI sessions
A day after 0.155.0's voice conversations and Touch ID, OpenAI patched a regression where reasoning summaries defaulted on for local TUI sessions and got requests rejected by providers that don't support them. Upgrade if you run Codex against a non-OpenAI backend.