Qwen3.8-27B-Perplexing-Pegasus-Heretic: One 27B Model for Local Vision, Tool Calling, and Typed Decisions
Running autonomous browser agents locally just got a lot simpler.
When we build and test WebBrain on local LLMs, peak agent performance usually comes from pairing two or three separate models in parallel:
- A vision-capable LLM to see and reason over UI and browser state.
- A fast Jev-style / typed-decision model for structured routing and scoring.
- Uncensored (abliterated) alignment so real-world agent workflows don't stall on false-positive refusals.
On local hardware, juggling multiple 20B+ models at once is brutal on VRAM and latency.
So we built and released one unified model suite that replaces the multi-model stack on a single consumer GPU: Qwen3.8-27B-Perplexing-Pegasus-Heretic.
What it packs into a single 27B checkpoint
- Multimodal Vision + Chat + Thinking (
<think>) + Tool Calling: Built on Qwen3.8-27B, supporting screenshot understanding, reasoning traces, and native tool calls in one model. - Perplexity
pplx-decidertyped-decision readout: Preserves thedecision/head alongside an intactlm_headfor structured routing and scoring. - Heretic ARA abliteration: Achieves 0/8 benchmark refusal probes while maintaining 100% (30/30) typed-decision agreement against the original Pegasus model.
- Multi-Token Prediction (MTP): All 15
mtp.*tensors are preserved, withblk.64kept atQ8_0in IQ/3-bit GGUFs. - Single-GPU local fit: Runs in 16GB VRAM (
IQ3_XXS/Q3_K_M+mmproj) or 24GB VRAM (Q4_K_M+mmproj).
Four hardware-ready Hugging Face repositories (Apache-2.0)
We published the full open-source release under Apache-2.0 across four Hugging Face repositories under WebBrain:
| Format / Target | Runtime & Hardware | Repository |
|---|---|---|
| imatrix GGUF + mmproj F16 | llama.cpp / LM Studio (16GB–24GB GPUs) | webbrain-one/Qwen3.8-27B-Perplexing-Pegasus-Heretic-GGUF |
| AutoRound INT4 | vLLM (Intel Arc Pro B70 32GB XPU & NVIDIA CUDA) | webbrain-one/Qwen3.8-27B-Perplexing-Pegasus-Heretic-AutoRound-INT4 |
| NVFP4 + FP8 | vLLM (NVIDIA Blackwell sm_100/sm_120 with MTP) | webbrain-one/Qwen3.8-27B-Perplexing-Pegasus-Heretic-NVFP4 |
| Base BF16 weights | Full-precision reference & downstream quantization | webbrain-one/Qwen3.8-27B-Perplexing-Pegasus-Heretic |
Acknowledgments
Big thanks to the Alibaba Qwen team, Perplexity, and AveLabs (AvecdrA) on Hugging Face for the foundational models this builds on.
If you're building WebBrain workflows or any local vision + decision agent stack, give it a spin and let us know what hardware you're running it on.
Tags: #OpenSourceAI #LocalLLM #AIAgents #WebBrain #vLLM #llamacpp #Qwen #ComputerUse #IntelArc #Blackwell