> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# NanoJev opens a Jev-inspired decision model for local testing
- URL: https://globalfeed.ai/en/nanojev-opens-a-jev-inspired-decision-model-for-local-testing/
- Published: 2026-09-19T22:34:22.000Z
- Updated: 2026-09-19T22:34:22.000Z
- Description: NanoJev publishes code, weights and training data for a Qwen3-0.6B-based decision model. Its game demonstrations also depend on shared planning code.
- Author: GlobalFeed Editor
- Tags: NanoJev, Jev, local ai, x-typesafeai, dil-en, elle, video

[TianyuCodings' NanoJev project](https://github.com/TianyuCodings/NanoJev?ref=globalfeed.ai) makes a small decision model available for local experiments, assigning probabilities to supplied options rather than writing long answers. Code, downloadable weights and training data have been released together. This is a separate Jev-inspired implementation, not the release of TypeSafe's own Jev model.

According to its [model card](https://huggingface.co/C-Tianyu/NanoJev?ref=globalfeed.ai), NanoJev adds decision heads to Qwen3-0.6B. Inputs describe a state, a question and the available options. Outputs are candidate probability distributions, Boolean assessments or ordered scores, rather than free-form text. Multiple questions can be evaluated in the same computation batch. Defining the options that an application offers remains the developer's responsibility.

The [published data package](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data?ref=globalfeed.ai) includes training and evaluation splits, frozen game scenarios and recorded action sequences. File hashes and verification scripts are provided as well. This makes it possible to inspect the training inputs and the steps behind a demonstration instead of relying only on its final result.

## The demonstrations include a planner, not just a model

The [game reproduction guide](https://github.com/TianyuCodings/NanoJev/blob/main/docs/GAME%5FRELEASE.md?ref=globalfeed.ai) identifies separate trained checkpoints for the maze and Snake. In the maze, code explores untried routes; in Snake, it filters immediate collisions and computes paths toward food. The model makes decisions within that system. The footage therefore does not establish that one general-purpose model learned the games unaided or outperforms Jev across tasks. These are the project team's experiments, not measurements reproduced by GlobalFeed.

The [pipeline guide](https://github.com/TianyuCodings/NanoJev/blob/main/research/pipeline%5Frunbook.md?ref=globalfeed.ai) calls for a compatible NVIDIA CUDA environment. Its programmatically supervised training workflow does not require a model-service API key, but the published game runs were recorded on A100 GPUs. A small backbone alone does not establish speed or reliability on an ordinary computer. The release enables developers to test this decision approach on their own data; a correctly formatted probability is not a guarantee of a correct decision.