cu-Jev
CUDA-native engine that serves open Qwen3.5 checkpoints behind a Jev-compatible API, reading option logits without decoding.
Overview
cu-Jev is a decision inference engine written in C and CUDA that exposes the same wire format as TypeSafe's Jev: POST /v1/systemone with a state and typed questions (noul, choice, score) returns typed answers with probabilities and confidence. It turns open Qwen3.5 checkpoints from 0.8B to 9B into decision functions, with no PyTorch at runtime.
The state runs once and each question is a short isolated branch that attends to the shared KV cache and recurrent state; at the last token the engine reads the logits of single-token option labels directly from the LM head. Pointing TYPESAFE_BASE_URL at cu-Jev lets the official SDKs work unchanged. The repo ships a Starfighter browser game in which the model answers four typed questions per tick (move, fire, shield, threat); the README reports about 95 ms per decision for Qwen3.5-4B on an RTX 3090, and describes validation against transformers and reproducible benchmarks.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | move: stay or step to a neighbouring lane? | — |
| Yes / no | fire / shield: should the ship fire or raise its shield now? | yes · no |
| Score | threat: how dangerous is the current situation? | — |
How it works
- Load a Qwen3.5 checkpoint into the CUDA engine
- Receive a Jev-format request; run the state once
- Branch each question off the shared state and read option logits
- Return answers with probabilities and confidence in Jev's format
Strengths
- Self-hosted, Jev-compatible API usable with official SDKs
- Shared-state branching avoids recomputing the state per question
- No text decoding; hand-written kernels
Caveats
- Requires an NVIDIA GPU
- Accuracy depends on the open base checkpoint, not Jev's trained model
Getting started
uv run cujev serve --model models/Qwen3.5-4B, then open http://127.0.0.1:8080/play for the Starfighter demo.At a glance
- Type
- Benchmark
- Status
- beta
- Language
- C/CUDA
- Runtime
- local HTTP server
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Inference engine: a CUDA-native implementation of the Jev System One API that keeps decisions GPU-resident, shipping a Starfighter demo and a benchmark script.
yibie/awesome-jevCuda-Jev — a CUDA-native Jev System One decision inference engine. Jev compatible API, examples, and reproducible benchmarks.
hellogumbo/awesome-jev
Frequently asked questions
Is cu-Jev the real Jev model?
No. It serves open Qwen3.5 checkpoints using Jev's API format.
Can I use TypeSafe's SDKs with it?
Yes, point TYPESAFE_BASE_URL at the cu-Jev server.
How fast is it?
The README reports about 95 ms per decision for Qwen3.5-4B on an RTX 3090 in the Starfighter demo.
Related projects
Docker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.
ServicebetaPythonCore ML and Neural Engine port of Laya for Apple Silicon, about 5 ms per short decision on an M3 Max with energy benchmarks.
BenchmarkbetaPythonToy local Jev-shaped decision API on a 0.5B model that answers yes/no, choice and score questions by scoring option tokens.
ServiceprototypePythonSmall open decision models (0.8B and 4B, from Qwen3.5) with a server that speaks Jev's /v1/systemone format in one forward pass.
ServicebetaPythonServes MiniCPM5-2B-8bit on Apple Silicon as a Jev-compatible /v1/systemone service and reproduces its JevBench score.
ModelprototypePython- reflex★ 166
Small open decision model on a frozen Qwen3.5-4B serving /v1/systemone in about 200 ms, benchmarked against Jev on JevBench.
ServiceresearchPython