Unified 27B local agent model open-sourced: vision, typed decider, abliterated, runs on 16GB VRAM

ButtercupLyn100 · reddit · 2026-11-16

A Reddit user released Qwen3.8-27B-Perplexing-Pegasus-Heretic (Apache-2.0), a merged open-source checkpoint that unifies what local agent stacks usually run as three separate 20B+ models: a vision parser, a fast typed-decision model, and an abliterated model to avoid false-positive refusals.

Key details:

Four repos on Hugging Face under webbrain-one: GGUF + mmproj F16 (llama.cpp/LM Studio), AutoRound INT4 (vLLM on Intel Arc Pro B70 and CUDA), NVFP4 + FP8 (Blackwell sm100/sm120 with MTP), and BF16 base weights.

Original post →

More from coding & agent

coding & agent channel →