prithivMLmods/JEV-27B-GGUF

JEV-27B-GGUF

autotrust/JEV-27B is AutoTrust AI's second-generation integrated System 1 + System 2 open model, built on a frozen, bit-identical Qwen3.8-27B backbone with the "Blocks of Experts" recipe. System 2 is ordinary text generation and reasoning through the untouched base lm_head (78.0% HumanEval pass@1, with all 164 completions byte-identical to the base model). System 1 is a small, detachable 108.9M-parameter LoRA plus a 24-slot fp32 decision head that answers typed noul (yes/no), choice (2-16 options), and score (0-5 scale) questions in a single forward pass, distilled from the closed, hosted TypeSafe Jev 1.13's own output distributions via the Apache-2.0 SargeDev/jev-distill-corpus-v3 corpus. On 25,376 Jev-labelled held-out rows it reaches a mean KL of about 0.017 from the teacher (about 60 sampled decisions to gather one nat of evidence, with the teacher's mistakes reproduced too), 90.5% choice top-1 agreement, 0.995 noul AUROC, and ECE of 0.0009 with no post-hoc correction. It also transfers better to unseen task families than JEV-9B (OOD KL 0.104 vs 0.234), reaches 96% of the teacher's accuracy at 16 options on an independent human-labelled benchmark, and in AutoTrust's own runs edges the hosted Jev on a six-benchmark public mean (84.07 vs 83.85), scoring higher on JevBench, OpenJev text, Nimble and MASSIVE-en and lower on Kev and VitaminC. A single decision takes a median 137 ms on one B200, versus 238-301 ms independently measured for the hosted API, and one GPU sustains about 6x the benchmark throughput. Both systems are served from one vLLM engine with per-request routing, while the smaller JEV-9B remains the faster option. It is released under Apache-2.0 as an independent student with no shared weights, code, or affiliation with TypeSafe AI.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp