---

language:

• en
• zh
• ja
• ko
• fr
• de
• es
• pt
• it
• ru

license: apache-2.0

base_model:

• Qwen/Qwen3-VL-235B-A22B-Instruct

tags:

• text-generation
• vision
• multimodal
• zenlm
• zen
• gguf
• abliterated
• moe
• ocr
• document-understanding
• hanzo

pipeline_tag: text-generation

library_name: gguf

---

Zen Designer GGUF: 235B Vision-Language Model (Abliterated)

235B MoE | Vision-Language | GGUF Quantized | Abliterated

Fine-tuned from Qwen3-VL-235B-A22B-Instruct (Apache-2.0) with Hanzo identity + agentic-data training + abliteration, then GGUF-quantized. A 235B-total / 22B-active Mixture-of-Experts vision-language model supporting images, video, documents, charts, GUIs, and spatial reasoning with 256K context.

---

Model Specifications

| Attribute | Value |

|-----------|-------|

| Base Model | Qwen3-VL-235B-A22B-Instruct (Apache-2.0) |

| Parameters | 235B total / 22B active (MoE) |

| Architecture | Vision-language transformer (Mixture of Experts) |

| Context Window | 256K tokens |

| Modalities | Text, Images, Video, Documents |

| OCR Languages | 32 scripts |

| License | Apache 2.0 |

---

Available Formats

| Format | Size | Description | Recommended Use |

|--------|------|-------------|-----------------|

| Q2_K (split) | ~60 GB | 2-bit quantization, 15-part split | Servers with 64+ GB RAM, maximum scale |

| Q4_K_M | ~142 GB | 4-bit quantization, single or split | Best quality/size tradeoff for local inference |

---

Quick Start

llama.cpp

# Download a split (Q2_K example — replace with Q4_K_M filename as appropriate)
# Then run:
llama-cli \
  --model zen-designer-235b-a22b-instruct-abliterated-Q2_K-00001-of-00015.gguf \
  --mmproj mmproj-zen-designer-235b-a22b-instruct-abliterated-f16.gguf \
  --image your_image.jpg \
  --prompt "Describe this image in detail." \
  -n 1024 \
  --ctx-size 8192 \
  --temp 0.7

For multi-part files, place all split parts in the same directory and point --model to part 00001.

Vision Tasks

Zen Designer handles a broad range of visual inputs:

• Image analysis and description
• Document and PDF parsing
• Chart and table extraction
• GUI navigation and screen understanding
• Video understanding with temporal reasoning
• Bounding box and spatial grounding

---

Abliteration

This model has been abliterated — a technique that removes refusal behaviors encoded in the model weights without fine-tuning. The process works by identifying the refusal direction in the model's residual stream and projecting it out of the weight matrices.

What abliteration does:

• Removes hardcoded refusal responses
• Preserves all other capabilities and knowledge
• Does not alter factual knowledge or reasoning ability

What abliteration does not do:

• Add harmful knowledge the base model lacked
• Guarantee any specific behavior
• Replace a system prompt or application-level safety policy

Users are responsible for appropriate deployment and use of abliterated models. Apply system prompts and application-layer controls to define model behavior for your use case.

---

Attribution

Built on Qwen3-VL-235B-A22B-Instruct by the Qwen team, Alibaba Group, released under the Apache License 2.0. Hanzo's contribution is identity training, agentic-data fine-tuning, and abliteration on top of that base, distributed here in GGUF format. The base model is used under the terms of the Apache License, Version 2.0.

---

Model Family

| Model | Format | Parameters | Context |

|-------|--------|-----------|---------|

| zen-designer-235b-a22b-instruct | SafeTensors | 235B / 22B active | 256K |

| zen-designer-gguf | GGUF | 235B / 22B active | 256K |

---

Links

Zen LM | Hanzo AI | GitHub | All Models

---

Part of the Zen model family (zenlm.org) by Hanzo AI (Techstars '17) and Zoo Labs Foundation (zoo.ngo).