Jev directory
CommunityJev Models & Reimplementations
openjev on Qwen3.5
Alex Nikolic's open reproduction of Jev as a Qwen3.5 cross-encoder: one NLI primitive over premise and hypothesis - entailment, contradiction or neutral - that reranks answers, grades them, guards content and even picks game moves, with the 4B v5 checkpoint near Jev 1.13 on public JevBench items.
- Reading time
- 3 min
- Published by
- Community
- Author
- AlexWortega
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceRuns locally
Highlights
- Qwen3.5 as an NLI cross-encoder with three labels - entailment, contradiction, neutral - for reranking, grading, content guards and game moves.
- JevBench v1.2 public items: 4B v5 scores 0.814 overall against Jev 1.13's 0.866 and SemIf's 0.810, with 0.622 on the hard tier.
- Option order does not matter: 0 label changes over 231 items asked four times, largest probability change 1.8e-7, bit-identical on repeats.
- Author-measured weak spots: 0.600 accuracy as a shell-command guard catching 37% of deny-worthy commands, and 0.833 to 0.467 under one adversarial line.
- v5 trained on train and test splits of MMLU, ARC, GSM8K, HellaSwag, WinoGrande, GPQA-diamond, CLINC-150, Banking77 and ESCI, so those scores are not reported.
Quickstart
python
from openjev_decide import OpenJev
jev = OpenJev.from_pretrained("AlexWortega/openjev", subfolder="qwen3.5-4b-nli-v5", device="cuda")
jev.decide(
"Policy: refunds require a receipt and purchase within 30 days. The customer bought 12 days ago but has no receipt.",
[{"type": "noul", "instructions": "Under the stated policy, is a refund permitted?", "options": ["no", "yes"]}],
)
# -> [{'noul': 0.07}] - the probability of "yes"Watch out
Copy reviewed against the model card fetched on 2026-10-15, truncated at 12,000 characters before the serving and reproduction sections. The card is candid about its own security weaknesses. MIT licensed with 648 likes at fetch time. The madewithjev listing calls it an MLP on top of Qwen 4B, while the card documents Qwen3.5 sequence-classification checkpoints.
Questions about this entry
Short answers drawn from the linked project or page.
Qwen3.5 turned into a Jev-style model: a cross-encoder that reads a premise and a hypothesis and answers with entailment, contradiction or neutral. From that single primitive it reranks answers, grades them against a reference, guards content and chooses game moves, per the model card.
Typed decisions. Hand it a state and a closed set of options with rubrics and it returns a probability for each option in one forward pass per option, using the model's own softmax; the card says the answer does not depend on the order the options come in.
On JevBench v1.2 public items, 4B v5 scores 0.814 across all 231 items against Jev 1.13's 0.866 and SemIf's 0.810; on the hard tier it scores 0.622 versus Jev's 0.730. The judge tier is held out by the benchmark, so no JevBench Score is claimed.
The author measured 0.600 accuracy as a shell-command safety reviewer, catching only 37% of deny-worthy commands, and one adversarial line in the state dropped JevBench accuracy from 0.833 to 0.467 and raised allowed deny-worthy commands from 17% to 85%.
The card reports a test over all 231 public items asked four times: 0 label changes for repeats, reversed and shuffled options, with a largest probability change of 1.8e-7 and bit-identical output on repeated identical requests.
Weights load with transformers (AutoModelForSequenceClassification, subfolder qwen3.5-4b-nli-v5) or through the OpenJev.decide helper in code/openjev_decide.py, and code/sglang_openjev serves it through SGLang at 1.5-3x the throughput of the transformers path.
Because v5 deliberately includes the train and test splits of MMLU, ARC, GSM8K, HellaSwag, WinoGrande, GPQA-diamond, CLINC-150, Banking77 and ESCI in its mixture. Numbers on those benchmarks are not reported; no JevBench item and nothing from the reported tables' benchmarks was ever in the mixture.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Gestalt Labs' Jeff is a LoRA adapter on Qwen3-4B-Instruct-2507 that answers Jev-style choice, noul and score questions with labels and probabilities instead of prose, runs locally, and ships a train-your-own guide; on 9,730 fact-checking examples it trails Jev 1.13.0 on accuracy (0.8183 vs 0.8283).
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonIndependent training toolkit and open weights for System One decision models in the same category as Jev: 17M, 68M, and 400M models answering Choice, Noul, Score, and reranking questions, with Python and ONNX inference and a documented generalization gap to Jev 1.13 on the project's S1MB benchmark.
Jev Models & Reimplementations#community#decision-models#onnx
Community
hotchpotch/bekko-system-onePython
hotchpotch/bekko-system-onePythonAn open, local attempt to bring Jev-style typed decisions to images: a Qwen3-VL backbone returns numerical distributions for independent choice, score and noul questions without generating answer text.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
andrueandersoncs/visual-jevPython
andrueandersoncs/visual-jevPythonAn open-source, Jev-compatible System One model: Dex answers choice, score and yes/no questions over text in a single forward pass from an Ettin-1B encoder, and speaks the same /v1/systemone shape as Jev.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
canadaornaw/dexPython
canadaornaw/dexPythonA Q8_0 GGUF of knowlet's Gemma-4-12B-Unified-System-One with a native System One decision runtime and an independently fitted GGUF temperature, shipped with an F16 mmproj for the model's native image and audio projections.
Jev Models & Reimplementations#community#open-models#quantized
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.