whereismyai
NanoJev
NanoJev is a 0.6B parallel decision model that outputs complete probability distributions without token decoding.
NanoJev is a 0.6B parallel decision model based on Qwen3-0.6B that processes states and questions to output complete probability distributions without output-token decoding. It handles multiple tasks including ViZDoom Basic, ViZDoom Predict Position, Maze, and Snake using a shared checkpoint and decision heads. The system supports parallel decisions by batching independent states, questions, and candidate paths in one backbone forward pass. It provides dynamic candidates, boolean and ordered scores, and direct probabilities to rank, select, or sample actions. The product is designed for researchers and developers working on decision-making tasks, offering local inference services and a complete dataset and training pipeline.
NanoJev at a glance
- Best for: Software engineer, Researcher
- Runs on: CLI, API, open source
- Pricing: Free
- Released: September 2026
- Category: AI research tool
- Website: github.com/tianyucodings/nanojev
Key features
- Parallel decisions across independent states
- Dynamic candidates returning score distributions
- Boolean score prediction for propositions
- Ordered score predictions over multiple levels
- Direct action ranking, selection, and sampling
- Shared backbone using Qwen3-0.6B
Use it for
- Playing ViZDoom Basic and Predict Position tasks
- Navigating a 50 by 50 maze
- Collecting food items in Snake game runs
- Evaluating decision models with local replays
Alternatives to NanoJev
- SemIf OpenJev — Extracts direct token probabilities without generating text or running JSON repair loops.
- Jiwo — It provides low-latency decision-making with a 50ms median request time.
- Vev — Runs locally on NVIDIA hardware with a fast, Jev-compatible API.
- Laya — It runs typed decisions in a single forward pass over 100+ languages.
- Kev — Provides calibrated probabilities and matches the TypeSafe API.
- Vegaml — Runs locally on your hardware with calibrated probabilities and conformal answer sets.
- Decido — Runs locally on CPU while beating published benchmark scores.
- Nouli — It gives stateless decision models a persistent memory across branching sessions.
- ej — Runs locally on CPU from a single 11 MB file without APIs or GPUs.
- Deepopen — Provides millisecond-level inference speeds and completely eliminates text generation hallucinations.
- Clef — Offers low latency and high-speed classification for agentic workflows.
- Homi — It answers complex audio questions natively with high accuracy.
All NanoJev alternatives