Jev alternatives

NanoJev

The only open project that actually beats Jev at something

Quick answer

NanoJev wins where latency and repetition matter: 128/128 on ViZDoom Basic against Jev’s 56/128, on a 0.6B model you can run locally. But the four games it is evaluated on are the four games it was trained on, and Jev still wins the maze test set 7/10 to 4/10. Treat it as evidence that a small purpose-trained model can outrun a hosted decision API inside a tight loop — not as a general classifier.

NanoJev is a Qwen3-0.6B backbone with parallel decision heads, trained on 18,760 decision questions drawn from four game environments. It is the only open project in this comparison that wins a task outright against Jev — and the honest reading of that win is that the benchmark lives inside its own training distribution. Keep it for control loops, not for tickets.

Train a decision model

A small non-autoregressive encoder with decision heads. No text generation at all — the classic classifier shape, rebuilt for natural-language options.

Search aliases

nanojevNanoJevNano JevTianyuCodings/NanoJev

Key specs

Licence
Not stated in the repository; third parties list MIT
Author
TianyuCodings
Backbone
Qwen3-0.6B + parallel decision heads
Size
0.6B parameters
Latency
Inference script expects CUDA — no Apple Silicon path
Wire format
Own harness; not a /v1/systemone server
Install
See repository (Chinese README available)

NanoJev against Jev, task by task

These are the project’s own held-out numbers. Read them alongside the scope caveat: four games trained, four games evaluated.

NanoJev against Jev, task by taskNanoJevJev
ViZDoom Basic128/12856/128
Predict Position27/12811/128
50x50 maze, attempts needed2252,738
Maze test set4/107/10
Snake8/88/8 (tie)
ScopeTrained and evaluated on the same four gamesGeneral-purpose hosted model

When NanoJev is the right choice

Use NanoJev when your decisions live inside a tight control loop — a game agent, a simulator, a rapid branch evaluator — and you want the whole thing on your own GPU. It proves the useful point that a 0.6B purpose-trained model can beat a frontier decision API on a narrow, repeatedly-sampled task, which is exactly the regime where per-call latency and cost dominate.

When to walk away

Do not use NanoJev to triage tickets, moderate content or route support mail. Its numbers come from four game environments it was trained on, Jev still wins the maze test set, and an independent re-run of its maze harness showed that no LLM readout — not even a much larger one — beat the majority label on the simple question "is one step north clear?". There is also no Apple Silicon path, and the repository does not state a licence.

Adopting NanoJev in three steps

01Check the licence first — the repository does not state one, so a commercial deployment needs legal review before anything else.
02Run it on CUDA hardware and validate on your own environment. The published wins come from the four games it trained on.
03Instrument attempts and collisions, not just success rate: in the project’s own harness the attempt counts moved far more than the goal test did.
python / parallel decisions per step
# NanoJev answers several questions about one state in a single
# forward pass, which is what makes a control loop viable.
#
# Published harness results (held-out gameplay):
#   ViZDoom Basic      128/128   (Jev 56/128)
#   Predict Position    27/128   (Jev 11/128)
#   50x50 maze          225 attempts (Jev 2,738)
#   maze test set       4/10     (Jev 7/10)
#   Snake               8/8      (tie)
#
# Train/eval pipeline ships with the repo. The inference script
# expects CUDA. Evaluate in your own environment before trusting
# these numbers: the four games are also the training set.
The reusable idea is parallel decisions over dynamic candidate sets in one forward pass — that is what makes a per-step decision affordable.

NanoJev questions people actually ask

Does NanoJev really beat Jev?

On two of the tasks the project reports, yes: ViZDoom Basic 128/128 against 56/128 and Predict Position 27/128 against 11/128, plus far fewer maze attempts. On the maze test set Jev still wins 7/10 to 4/10, and Snake is a tie. The win is real but scoped to the four games the model was trained on.

Can I use NanoJev as a general classifier?

No. It is a 0.6B model trained on 18,760 decision questions from four game environments and evaluated on those same four. Nothing about that setup transfers to support triage, moderation or routing, and the project makes no such claim.

Is the license safe for commercial use?

Unclear. The repository does not state a licence — several third-party write-ups list MIT, but that is not the same as a licence file. Get legal review before shipping it internally, and do not rely on a blog post for the terms.

Are the benchmark numbers trustworthy?

They are honest but narrow. A separate project re-ran NanoJev’s own maze harness and found that the result depends heavily on how you read the model: the same untuned Qwen3-0.6B went from 13/15 mazes and 20,555 attempts under one readout to 15/15 and 5,825 under another. That is a property of the readout, not of map understanding.

Sources

Every comparison number on this page is a third-party published figure or a read of a public repository — not a benchmark we ran. We have not tested TypeSafe Jev itself, and its customer agreement forbids using its outputs to build similar products.