IdeaNJEV: An Open-Source 4B "System-1" Model That Outperforms Its Own Base at Decisions
Natural language in, calibrated JSON out, in one forward pass. Our open, MIT-licensed take on Jev: 77.9% accurate on the typed-decisions benchmark, trained on a single NVIDIA DGX Spark.
September 23, 2026 · 4 min read
On 15 September a lab called TypeSafe AI released a model that doesn't write. Jev takes a short description of a situation and a set of typed questions and returns one answer per question, with a calibrated confidence, in a single forward pass. It is closed source. At IdeaNirvana AI we built an open one: IdeaNJEV, a 4B "System-1" model on Qwen3.5-4B, released under the MIT license.
Why we cared
Most of what we build is multi-agent systems: a request arrives, something decides where it goes, whether it needs escalation, how risky it is, and only then does a capable model reason and write. That first decision is a triage step and does not need a model that thinks for a minute. A fast, private decision gate in front of the expensive model is directly useful to us.
What we measured
On the benchmark's full 2,000 test decisions: 77.9% accuracy and a calibration error (ECE) of 0.0154, in the same range as the other open Jev-style models released in the days since (Laya, Verdict). The gaps are about one standard error, so we call it "at the top of the range", not a win.
Against the original Qwen3.5-4B writing the same JSON on the same GPU (200 decisions), IdeaNJEV scored 73.5% against 49.5% and was about 3x faster (373 ms against 1,118 ms). With thinking on, 16 of 20 decisions had not finished within 3,000 tokens. It is not the fastest model in its class: a small encoder beat us on raw latency. The base model was prompted, while IdeaNJEV was trained on this benchmark, so the accuracy gap measures what the training bought.
Where we started
We began by training a model completely from scratch: Andrej Karpathy's nanochat, 561M parameters, 9.9 days on a single
NVIDIA DGX Spark. Turned into a decision model it scored exactly what random guessing scores (31.75%), and after fixing
three ways it had learned to cheat, 55%. So we kept the design and moved it onto a pretrained backbone, which took it to 78%.
Where this goes next
First as a standalone app, then, once it has earned it, as a classification gate inside one of our agent workflows.
Read the full story on Medium. Code, the side-by-side demo against Laya's hosted model, and every number are open source at github.com/bhaweshkrsingh/IdeaNJEV; the weights are on Hugging Face.
Thanks to Andrej Karpathy for nanochat, the Qwen team, TypeSafe AI, Convai Innovations (Laya), and everyone who published openly since.