NanoJev Tutorial: Install the 0.6B Open-Source Jev Replica and Watch It Route Decisions at 47 ms

This is Alex Hitt's six-and-a-half-minute local install of NanoJev, the open-source nano replica of the Jev System One architecture — GitHub TianyuCodings/NanoJev and Hugging Face C-Tianyu/NanoJev (write the full paths; there is an unrelated sdmlai/nano-jev on the Hub). The model is a 0.6B parallel decision machine: a Qwen3-0.6B backbone with decision heads and no vocabulary generation head, so it cannot emit conversational text even if you ask — it only estimates probabilities. Because the shared state is encoded exactly once and the questions are distributed in parallel, the GPU answers in a single matrix pass; the video's demo badges read 34-47 ms. The install: clone the repo and pip install -r requirements-toy.txt (there is no requirements.txt and no CPU variant — the three files are requirements-toy.txt, requirements-vizdoom.txt, requirements-shooting-demo.txt). The core file pins torch==2.14.0 (with transformers==5.17.0), and on newer hardware such as an RTX 5090 that default binary may not recognize the architecture — manually install a CUDA-compatible wheel or the model falls back to CPU and latency jumps from ~50 ms to over 400 ms, the video's own 'CPU Fallback Destroys Latency' chart. Weights ride separately from code: snapshot_download(repo_id="C-Tianyu/NanoJev", revision="unified-games-v1", local_dir="checkpoints/NanoJev-unified", allow_patterns=["best.safetensors", "config.json", "tokenizer/*", "backbone_config/*"]) — the video's on-screen pseudo-code writes repo_id="unified-games-v1", but in the real repo unified-games-v1 is the revision tag (the step-400 checkpoint of the hard_lr1e5 run, one model for Maze, Snake, ViZDoom Basic and Predict Position). Miss the best.safetensors pattern and you get an empty local directory; the video also warns about HF token gating, while the repo's current README says the release is public and downloadable without signing in — either way, an empty folder means a wrong pattern or missing authorization. Verification runs as two services: the PyTorch inference engine on 127.0.0.1:8765 (repo entry: scripts/serve_decisions.py, POST /api/evaluate) and an HTML front-end on 8080, where the dashboard shows NanoJev navigating a 50x50 maze — header 'Environment: 50x50 Maze Grid (OOD)' — with per-direction confidence bars (North 97.2%, South 2.9%, East 3.7%, West 3.2% in our captured frame) and a 47 ms System One latency badge. Then the routing script: point sys.path at the checkpoints folder, import the prediction class, and respect the entry contract — no conversational prompts. Build the shared state payload (application logs, code diffs, parsed DOM trees), define typed questions against it (categorical choice, boolean, ordered score), and one forward pass returns a nested dictionary of normalized floats — demo card: choice 0.842, boolean 0.991, score 0.740 — no strings, no JSON chain-parsing, $0.00 exit-token fees. The routing layer is the asymmetric-error principle: 85%+ executes autonomously, 50-85% hits the ambiguity gate and a human, below 50% reroutes to a backup frontier model — numerical gates that filter high-volume work without API fees. The honest ending, kept on purpose: the unified checkpoint is heavily conditioned by physical positioning and game spatial mechanics — the video's own chart shows ~88% on spatial navigation collapsing to near zero on semantic text — so before production, run your own supervised fine-tuning flow with the repo's monitoring scripts (the closing card shows Fine-Tuning Epoch 1/5). Local System One models give you autonomous apps that react at software speed; the fine-tune is what makes them yours.