Jev's Open-Source Clones Arrive in 48 Hours: 9B Nimble & 0.5B Kev

Two Open-Source Jev Alternatives Emerge Within 48 Hours

Jev, a System One decision model from TypeSafe that scores candidate answers without generating reasoning, went viral but remained closed-source. Within two days, two independent teams released fully open-source alternatives: Bespoke Nimble (9B parameters) from Bespoke Labs and Kev (0.5B–8B) from Jared Palmer. Neither project uses distillation from Jev; they prove the approach is reproducible without secret recipes.

Nimble: Contrastive Data Curation Flips Decisions with a Single Word Change

Nimble's core innovation is contrastive data curation . The method constructs pairs of nearly identical samples where only one critical fact changes, forcing the correct answer to flip. The model must learn which evidence actually drives the decision.

Concrete example: a policy states "Only Mira can authorize refunds for account 42." Two samples keep the policy text identical but change the signature on the authorization record from Mira to Noah. The first sample's answer is true, the second false. All other text is identical.

Benefits: no human probability annotation needed, no teacher-model distillation required. Labels are hard (correct/incorrect), and the model learns to discriminate on its own.

Training data covers 10 domains (Commerce, Education, Media, Public Services, Supply Chain, Travel, Workplace, etc.) and three task types: Choice, Boolean (Noul), and Score. Total training set: 2,676 samples.

Inference: Nimble does not generate chain-of-thought or JSON. It reads logits of candidate answer tokens, converts to probabilities, and assembles structured output programmatically — generating zero extra tokens.

Benchmarks on their 324-sample held-out test (162 contrastive pairs, 6 domains):

Base Qwen3.5-9B (untuned): 66.36%

Qwen3.8-27B (3× parameters): 84.88%

Nimble (9B LoRA): 90.12%

Jev (closed): 93.21%

A 9B model surpasses a 27B model by >5 points and trails Jev by only 3 points.

Kev: Smaller, Runs on a Laptop

Kev targets minimal viable deployment. Based on Qwen2.5-0.5B, it uses a LoRA adapter (8.8M params) plus a small readout head (0.46M) — only 9.3M trainable parameters (1.9% of backbone). Training on an M5 MacBook Pro takes 1 hour 45 minutes and consumes 0.06 kWh.

Architecture: the state is encoded once; multiple questions run independently under a block-causal mask so one question cannot see another's content. Isolation verified to 4e-6 precision.

Jared Palmer built a browser Playground with preset scenarios, live state/question editing, and even a chess demo where each move is a Choice question, the board is the state, and position evaluation is a Score question.

Kev releases four sizes: 0.5B, 0.6B, 4B, 8B. The recommended kev-4b balances accuracy and size. Out-of-domain scores: kev-4b 0.79, kev-8b 0.80, Jev 0.86. Gaps appear mainly on knowledge-heavy tasks (MMLU 0.69–0.75 vs. Jev's 0.90) and date arithmetic.

Key finding: learning rate controls catastrophic forgetting of base knowledge. Default LR 2e-4 erases pre-trained knowledge; lowering to 5e-5 preserves most of it. This suggests the fine-tuning teaches the model how to use existing capabilities for decision-making, not to learn new facts.

Barriers Lowering: No Distillation Needed, But Limitations Remain

Both projects explicitly avoid Jev outputs: Nimble states "only used Jev for evaluation, not for outputs"; Kev likewise. They demonstrate that System One decision models do not require proprietary recipes.

Limitations noted:

Nimble's test set is small (324 samples, 162 pairs, 6 domains). Authors admit "no standard benchmark; Nimble may perform far worse than Jev on other tests."

Kev's out-of-domain scores lag Jev by 6–7 points.

Community feedback: 90% single held-out accuracy can mask production risks; calibration curves and OOD tests are needed.

Critique: 9B is still large for embedding directly into programs like traditional classifiers.

Counterpoint: traditional classifiers cannot read natural-language policies and judge compliance — Nimble/Kev can.

The real signal is speed: Jev released, two independent teams delivered usable open alternatives in two days. Whether Jev is overhyped or not, the barrier for this class of model is dropping fast. Bespoke Labs' founder (previously OpenThoughts) says "Money likes speed" — this two-day turnaround proves it again.

Open-Source Links

Bespoke Nimble :

GitHub: https://github.com/bespokelabsai/nimble Hugging Face: https://huggingface.co/bespokelabs/Bespoke-Nimble-9B Kev :

GitHub: https://github.com/jaredpalmer/kev Hugging Face:

https://huggingface.co/collections/jaredpalmer/kev-6aad9d0ea49f2589665e07cd