Jev directory
CommunityJev Models & Reimplementations
Jev 9B Ft V3
A LoRA adapter for JEV-9B-MLX-NVFP4 that judges the shell commands an AI coding agent wants to run and decides whether each one can execute without asking the user. It is the local judge behind the omp command gate in Quentin-M/harness.
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Rank 16, alpha 32 fp32 LoRA on every quantized linear of the last 12 of 32 layers: 75 modules, 15.0 M parameters, applied unmerged on the NVFP4 base.
- Blocks exfiltration, remote shells and C2, malware and attacker tooling, fetch-and-run code, credential theft, evasion and disk-scale destruction.
- On 1,306 labeled commands it blocked 97.8% of dangerous and 65 of 66 critical commands while passing 92.5% of routine work.
- p50 latency 1.0 s and p95 1.3 s on an M5 Max; the quantized MLX weights are 60.2 MB.
- Trained on 6,391 commands from ShellRisk-Bench train, tomngdev/shell-safety-v1.1 (MIT) and synthetic rows; no bench commands were used for training.
Questions about this entry
Short answers drawn from the linked project or page.
Use the harness's System One adapter with mlx-vlm on Apple Silicon: uv run bench/judge/adapters/jev_systemone.py <JEV-9B-MLX-NVFP4 dir> 11436 <this repo dir> --model-id jev-9b-ft-v3. Ask the compact-v3 questions in bench/judge/gates/jev-9b-ft-v3.json and decide with that file's rule.
It is an MLX adapter meant for Apple Silicon, applied unmerged on top of the NVFP4-quantized JEV-9B base. The adapter file is 60.2 MB, and the reported p50 and p95 latencies of 1.0 s and 1.3 s were measured on an M5 Max.
The model repository is licensed apache-2.0. The training data sources keep their own licenses, including tomngdev/shell-safety-v1.1 under MIT, alongside ShellRisk-Bench train and synthetic rows.
No inference provider deploys this model, and the harness adapter runs it locally through mlx-vlm on Apple Silicon. Commands are judged on-device rather than sent to a hosted endpoint.
Thresholds were tuned for this exact adapter by 2-fold cross-validation grouped by state at a 2% false-allow budget. On the 1,306-command bench it blocked 65 of 66 critical commands and passed 92.5% of routine work, so some routine commands are still blocked.
It is published by memtalow on Hugging Face and used as the local judge behind the omp command gate in Quentin-M/harness. The fine-tuning recipe is documented in bench/judge/finetune.md of the harness repo.
It is a security-first gate for shell commands an AI coding agent wants to run, not a general-purpose decision model. It also depends on the JEV-9B-MLX-NVFP4 base and MLX on Apple Silicon, and its thresholds were tuned for this adapter specifically.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
X: Decision outputs in Grep
X: Standard One SH (StandardThinking)
More Jev resources like this
A from-scratch Jev-style decision model on a frozen Qwen2.5-3B backbone with a custom decision head and hand-rolled LoRA adapters. It returns typed action distributions for routing, safety gating, and RAG filtering, reporting 83.38% accuracy at 23.95 ms p50 latency.
Jev Models & Reimplementations#open-models#python#routing
Communityabirshhishirroy/mini-jev
Patronus Studio's 151M-parameter Jev-like decision model for AI-agent security: given a serialized agent scene, a security question and answer options it returns calibrated probabilities in one pass, reporting 92.8 balanced accuracy on 10,576 held-out decisions.
Jev Models & Reimplementations#community#open-models#security
Community
A Qwen2.5-1.5B-Instruct fine-tune that emits typed dependency verdicts, safe, caution, or block with a recommendation, for Dependabot and Renovate pipelines, reporting 95.7% F1 on the card's 4,800-update OSV-Bench.
Jev Models & Reimplementations#community#open-models#local
Community
A 0.6B non-autoregressive decision model on Qwen3-0.6B that returns calibrated Choice, Score and Noul distributions in one forward pass, scoring 95.8% on its 22,110-item held-out test split.
Jev Models & Reimplementations#community#open-models#local
Community
Superagent's security triage model: a continual fine-tune of AutoJev-27B that returns calibrated probabilities for prompt-injection and content checks, with a release policy at 0.70 unsafe.
Jev Models & Reimplementations#community#open-models#local
Community
A native MLX runtime for Laya's open typed-decision weights: short decisions on Apple Silicon, with no generated tokens, PyTorch, or cloud API.
Jev Models & Reimplementations#community#python#open-models
Community
mizorewww/laya-mlxPython
mizorewww/laya-mlxPythonFrom 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.