---
datasets:
- SargeDev/jev-distill-corpus-v3
base_model:
- autotrust/JEV-27B
pipeline_tag: text-classification
library_name: transformers
tags:
- text-generation-inference
- llama-cpp
- system-one
- system-two
- blocks-of-experts
- typed-decisions
- decision-model
- calibrated-probabilities
- knowledge-distillation
- jev
- noul
- choice
- score
- qwen3_5
- qwen3_5_text
- text-generation
- dual-head
license: apache-2.0
language:
- en
---

# **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

| File Name | Quant Type | File Size | File Link | Description |
|-----------|------------|-----------|-----------|-------------|
| JEV-27B.BF16.gguf | BF16 | 53.8 GB | [Link](https://huggingface.co/prithivMLmods/JEV-27B-GGUF/blob/main/JEV-27B.BF16.gguf) | Full BF16 weights. Highest quality, largest file size. |
| JEV-27B.Q3_K_M.gguf | Q3_K_M | 13.3 GB | [Link](https://huggingface.co/prithivMLmods/JEV-27B-GGUF/blob/main/JEV-27B.Q3_K_M.gguf) | Low quality. |
| JEV-27B.Q4_K_M.gguf | Q4_K_M | 16.5 GB | [Link](https://huggingface.co/prithivMLmods/JEV-27B-GGUF/blob/main/JEV-27B.Q4_K_M.gguf) | Good quality, default size for most use cases, *recommended*. |
| JEV-27B.Q5_K_M.gguf | Q5_K_M | 19.2 GB | [Link](https://huggingface.co/prithivMLmods/JEV-27B-GGUF/blob/main/JEV-27B.Q5_K_M.gguf) | High quality, *recommended*. |

## llama.cpp

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