Jev Vllm

A 4-bit AWQ plus GPTQ build of AutoTrust's JEV-35B from AI Outfitters: INT4 weights with FP8 and INT8 activations, calibrated on all 256 experts, and served by vLLM on a single 24 GB GPU.

Typed decisions on a fine-tuned DiffusionGemma 26B-A4B: super-decider answers choice, noul (yes/no) and score questions with calibrated distributions in one call, speaks the Jev/TypeSafe System One request format and OpenAI chat completions, and runs on one NVIDIA L40S. NVFP4 weights are 18 GB.

A lightweight Jev/System One compatible HTTP gateway that turns an existing SGLang or vLLM deployment into a decision service by scoring candidate tokens: no training, no weight changes, noul/choice/score in and probabilities out.