NanoJev

How we reviewed this: NanoJev’s inference and training code requires a CUDA GPU, which our test machine doesn’t have, so we didn’t run the model. We read the code and documentation at commit 71a513b, ran its offline test suite (92 tests, all passing except 2 skipped for missing optional data), opened its local replay viewer, and inspected the published model and dataset files on Hugging Face.

What it does

NanoJev is an open “nano replica” of Jev: Qwen3-0.6B with added decision heads that return a full probability distribution for each question in one forward pass, with no generated text. It supports the same three question shapes as Jev:

• Choice over 2 to 255 candidates, using a shared scoring head with attention across the candidate set,
• Boolean questions that return the probability a statement is true, and
• Score over 2 to 10 ordered levels, returning the distribution and expected value.

Several states and questions can be batched into one pass, and a persistent service exposes it at POST /api/evaluate. The repository also includes maze and Snake environments that combine the model’s local judgments with ordinary planning code, and experiments with calibration objectives that the author describes as an implementation of RLCD. TypeSafe hasn’t published how its RLCD works, so treat that label as the author’s interpretation.

Setup

The weights (C-Tianyu/NanoJev) and data (C-Tianyu/NanoJev-Data) are public. Running the model needs a CUDA environment; the recorded training environment was an A100. The inference code refuses to start without CUDA.

You don’t need a GPU to look around. The recorded game replays run from a plain static web server:

python3 -m http.server 8080 --bind 127.0.0.1 --directory web

We opened the side-by-side and arcade replay pages this way. Both loaded without errors and made no requests to outside websites. The README and the replays include recorded comparisons with Jev and with the original Qwen3-0.6B. Consistent with our policy, we don’t reproduce the Jev figures here.

Documentation and tests

The documentation is unusually thorough, in both English and Chinese. There are pipeline runbooks, a detailed audit of how its inputs map to TypeSafe’s question contract, design reviews, hashed data splits kept separate by map, and replay-verification manifests. The offline suite covers the question contract, planning, and training objectives, and it passed on CPU.

Training data and licensing

Read this section before reusing the weights or data.

• The pipeline can use Jev as a teacher. scripts/teachers.mjs has two labeling routes: one that asks Jev for its probability distributions, and one that asks a generative LLM for labels.
• One published variant was trained on Jev’s labels. The dataset’s scaled_games_labeled_v4 records mark their teacher labels with the source jev_native_evaluation, and the published games_api_seed17 variant was trained with a teacher objective on that data. The dataset manifest describes these as “authorized teacher distributions”.
• The other variants use programmatic targets, from a teacher-trained start. The maze and Snake showcase variants (local_atomic_seed17, games_gold_seed17) and the calibration experiments train on ground truth computed by code or on simulated outcomes. All of them, however, continue from the same earlier checkpoint, v3_teacher_coords_multi_seed17. The repository doesn’t clearly state which teacher produced that checkpoint’s labels.
• Why it matters. TypeSafe’s customer agreement restricts using its outputs to train models that imitate the service. The project states its teacher data was authorized; we have no way to verify that. If you plan to build on these weights or this data, especially commercially, check the terms yourself.
• No license is declared on Hugging Face. The GitHub code is MIT licensed, but neither the model nor the dataset repository declares a license, and the base model’s own terms still apply.

Maintenance

NanoJev was published on September 17 in a single day of commits and hasn’t changed since. There’s one open issue. It reads as a research release rather than a maintained library.

Verdict

NanoJev is a remarkably complete attempt to rebuild the whole System One pipeline in the open, and its documentation alone is worth reading if you work on small decision models. To actually run it, you’ll need an NVIDIA GPU. Before reusing the weights or data, look carefully at where the training labels came from and at the missing license on Hugging Face.

For an open reproduction that reads option scores from an unmodified model, see SemIf. For a tiny scorer you can train on a laptop, see Jevlike. Our Jev alternatives guide compares these approaches.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.