NanoJev
NanoJev (github.
Overview
NanoJev (github.com/TianyuCodings/NanoJev) is Tianyu Codings' 0.6B Qwen3 backbone with shared decision heads for dynamic Choice (2 to 255 candidates), boolean propositions, and ordered Score levels. Each request supplies state, question, and candidates; one forward pass returns full probability vectors with zero output-token decoding. The unified-games-v1 Hugging Face release trains Maze, Snake, ViZDoom Basic, and ViZDoom Predict Position together; README tables compare held-out success rates against TypeSafe Jev and untuned Qwen3-0.6B on the same controllers. Public side-by-side replays and ViZDoom players live on nanojev-dev.tianyuchen99.chatgpt.site; model C-Tianyu/NanoJev and dataset C-Tianyu/NanoJev-Data pin revision unified-games-v1.
Problem: Researchers want System One shaped decision heads on tiny backbones without paying hosted inference or decoding answer tokens from a chat model.
Built for: ML engineers reproducing parallel Choice, boolean, and Score heads who need a unified checkpoint, public dataset, and replayable game harnesses.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- Tianyu Codings
- Handle
- @TianyuCodings
How Jev is used
- Role in the product flow
- Parallel game and simulation decisions from a single small checkpoint with shared Choice, boolean, and Score heads
- Primitives
- ChoiceScoreNoul
- State in
- Per-step game observations encoded as text state plus runtime-defined questions and candidate action paths documented in NanoJev harnesses and dataset rows.
- Decision out
- Softmax over supplied candidates for Choice, sigmoid probability for boolean items, and weighted level distribution for Score without generating answer tokens.
- Encode candidate paths through the Qwen3-0.6B backbone
- Apply shared decision heads for each question type in the batch
- Return probabilities for epsilon-greedy or argmax controllers
- Replay trajectories through independent simulators for evaluation
NanoJev is an open weight replica path, not a hosted classifier.dev substitute. Where SemIf targets general System One HTTP on models you pick, and Simple Jev wraps HF logits behind /v1/classifier, NanoJev trains decision heads directly on game-scale data and publishes the full pipeline plus HF weights. Agent builders can still steal the interface lesson: pack state once, ask many typed questions, threshold probabilities in code. Game demos are evidence that 0.6B parallel heads can beat untuned backbones and match or exceed published Jev scores on specific ViZDoom splits; treat leaderboard cells as research baselines until you rerun scripts locally.
Sourced performance claims
- README reports NanoJev 128/128 ViZDoom Basic test successes versus 56/128 for TypeSafe Jev on the matched harness.Source: github.com/TianyuCodings/NanoJev README
- Unified dataset ships 18,760 decision questions per target variant including 16,333 ViZDoom questions.Source: github.com/TianyuCodings/NanoJev README
- Public GitHub repo TianyuCodings/NanoJev had about two thousand one hundred eight stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- Unified-games-v1 checkpoint across four game tasks
- Hugging Face model and dataset with documented training mix weights
- Browser replays comparing NanoJev, TypeSafe Jev, and untuned Qwen
- End-to-end training and evaluation docs for Predict Position
Stack
- Python
- PyTorch
- Qwen3-0.6B
- Hugging Face Hub
- ViZDoom
Pricing: MIT licensed open source; you pay for GPUs, Hugging Face bandwidth, and any cloud you rent for training or inference.
Links
FAQ
- Is NanoJev the same as TypeSafe Jev?
- No. It is an independent 0.6B research replica with public weights and game benchmarks. TypeSafe Jev remains the closed hosted System One model.
- How does it differ from SemIf or Simple Jev?
- SemIf serves general open models with a System One shaped server. Simple Jev builds classifier HTTP from HF logits. NanoJev fine-tunes dedicated decision heads on a unified games dataset and publishes the training recipe.
- Which Hugging Face revision should I download?
- Use revision unified-games-v1 on C-Tianyu/NanoJev and C-Tianyu/NanoJev-Data so weights and labels match the README demos.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.