Research
Jev is closed-weight with no paper. The community reverse-engineered the inference shape in days — but the RLCD calibration training remains the real moat. Below: reproductions, and the independent evaluations that test the official claims.
Open-source reproductions (training pipelines)
| Project | ★ | Why it's worth your time |
|---|---|---|
| TheoLeeCJ/SemIf | 2.1k | Zero-training "semantic if" on an RTX 3090: read option logits from Qwen3.5-4B. 20 decisions/s with shared-prefix reuse; 0.845 agreement vs Jev's 0.883 on the public subset |
| vinnylarouge/jevlike | 1k | The first viral reimplementation; options-as-queries attention pooling. Explicitly not an RLCD reproduction |
| TianyuCodings/NanoJev | 1.2k | 0.6B end-to-end training pipeline with paired proper-reward learning — closest public thing to RLCD. Chinese-friendly docs, weights + data on HF |
| jaredpalmer/kev | 692 | API-compatible with /v1/systemone. The money stat: −1.8pp in-distribution but −19.1pp OOD vs real Jev — proof calibration training is the moat |
| arnabgho/rlcd-lite | 1 | GRPO + Brier reward reimplementation; key ablation: proper-scoring-rule rewards produce calibration, binary rewards destroy it |
| DECRUX9812/openjev-lm | 1 | Cheapest entry point: LoRA on CPU in 89 min at $0, 92.9% answer agreement with the API |
More: LightJev · jev-visual · razorback16/openjev · typed-decisions · openJev-verdict-2.0
Inference-layer reimplementations (no training)
| Project | ★ | What it does |
|---|---|---|
| ekzhang/openjev-sglang | 206 | Qwen3.6-35B-A3B + SGLang, prefill-only with radix cache, full official HTTP API |
| AlexWortega/openjev | — | First open weights; Qwen3.5-4B as a 3-class NLI cross-encoder, zero-shot Doom |
| bnsd55/jevmlx | 40 | Apple Silicon MLX |
| r-ms/mini-jev | 26 | Frozen Qwen3-4B, option-letter logits |
| Trecto34/openjev-fighting-ring | 0 | Puts Diffusion/RLCD/Block-Causal/NLI approaches in one arena |
Large-scale / pre-registered evaluations
| Evaluation | Finding |
|---|---|
| willkelly/jev-evaluation | 9 adversarial experiments, 123,805 requests, $12.69: calibration ECE 0.075 in-domain but fails OOD (random 3-SAT); polite authority injection moves 147/200 answers; batching 255 questions is genuinely free |
| exs-brady/jev-eval-evidence | 55,347 typed judgments, Jev vs 10 LLMs + keyword/regex floors, pre-registered with a blind cross-family replication |
| ickma2311/jev-baselines-eval | Banking77/CLINC150: Jev 0.832/0.870, but a 9 ms bge-small + logistic-regression baseline hits 0.933; measured latency only 2.2× vs a nano-LLM — classical baselines still win some lanes |
| Fox-Islam/jev-bias-bench | Counterfactual bias bench, 11,984 calls: 29 persona attributes across hiring/lending/clinical/bail scenarios, with noise floors and negative controls |
| kokuren333/jev-jmle-benchmark | 9 years of Japanese Medical Licensing Exams (3,556 items): 88.58%; companion manuscript on medRxiv |
| scienthoon/jev-ood-calibration | OpenBookQA 94.2% / ECE 0.024, but on an unknowable rule-based task 44.7% acc with stated p 0.74 — miscalibration sign flips by question type |
| EmilLindfors/jev-horingssvar-eval | Norwegian hearing documents: equal accuracy to DeepSeek V4.1 Flash at $0.22 vs $3.08 / 1k docs, 0.32 s vs 26 s median |
Focused task evals
| Evaluation | Finding |
|---|---|
| anisselbd/jev-phishing-bench | Naive question: 62.6% (regex: 91.8%). Atomic decomposition + logistic regression: 95.0% |
| bitnovus/jev-spam-eval | 98.3% on 18.5k emails ≈ TF-IDF, but 97.3% vs 72.5% under distribution drift — drift robustness is the real edge |
| mahlernim/jev-korean-benchmark | Reading comprehension nearly lossless; input reordering changes 13–14% of answers |
| rorshopping/jev-on-a-laptop | A 1.5B model can't write 28-field JSON but makes 28 schema-valid decisions in 0.4s; 7B is the sweet spot |
| jev-poker | Only 63% agreement with a solver, inverted confidence — feed it pre-digested intermediate conclusions |
More: jev-banking77-experiment · jev-vs-open-decision-models · mnist-text-input · jev-ja-eval · jev-decision-benchmarks · does-jev-confidence-mean-anything · decision-model-benchmark · jev-capability-atlas
Deep dives & reports
| Piece | Why read it |
|---|---|
| sgnt.ai — "You could have built Jev" | Best technical explainer; reconstructs the single-token-logit recipe and cross-tabulates the reproduction results |
| HackSing/jev-report | 52-page independent research report (中文) |
| mizzlelover/jev-hub | Aggregator of 714 X posts (226 demo videos, 114 long-form) from launch week — the entry point to the video firehose |
| 掘金：发布 3 天登顶 HN，我把 Jev 的源码和黑料都扒了一遍 | Vercel AI SDK source dive (中文) |
| lindfors.no — A first look at Jev | Careful first-principles cost/latency measurement on real documents |
| ic.work：TypeSafe 发布 Jev 模型 | Skeptical take (中文) |
More: ickas.dev · southbridge.ai · classmethod · mikulskibartosz.name
Other awesome lists
- Anil-matcha/awesome-jev-by-typesafe — the largest directory; evidence-backed submission rules
- yibie/awesome-jev — the strictest curation, explicit inclusion criteria
- yzfly/awesome-jev-zh — 中文, updated daily, honest about negative results
- AbdelStark/awesome-typesafe — the whole TypeSafe ecosystem
- cobanov/awesome-jev — public review notes and evidence sources
- fatwang2/awesome-jev — submissions reviewed by Jev itself