What it is
cu-Jev is a CUDA-native inference engine that turns Qwen3.5 checkpoints into a decision function. It provides a Jev-compatible API for sending a state and typed questions, returning probabilities and confidence. For developers needing fast, GPU-resident decision inference.
How it uses Jev
cu-Jev implements the Jev System One API: it accepts a state and typed questions (noul, choice, score), runs the state once, then evaluates each question as a short branch attending to the state's KV cache. It reads logits of option labels to produce calibrated probabilities and confidence, returned in the response.
Primitives:choicescorenoul
Technique worth stealing
Shared state with isolated branches: state KV cache and recurrent state are computed once and reused read-only across all questions.
Try it
uv run cujev serve --model models/Qwen3.5-4B --port 8080
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.95
- Shows a System One patternyes0.73
- Handles uncertaintyno0.04
- Measuredno0.09
- Runnableyes0.89
- Worth recommendingunclear0.40
- Model replicayes0.92
- Problem scopescore on a 0–2 scale0.99
- About Jevyes0.98
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.