GGUF version of insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated, with both MTP (Multi-Token Prediction) speculative decoding layers and full vision weights preserved.

Why this exists

Two problems with existing GGUF conversions of this model:

Missing MTP layers: The original HF→GGUF conversion using llama.cpp's convert_hf_to_gguf.py produced a broken GGUF — the metadata declared 33 layers (32 transformer + 1 MTP) but only 32 layers' tensors were actually exported, causing check_tensor_dims: tensor 'blk.32.attn_norm.weight' not found errors in LM Studio and other loaders.
Missing vision: insraq's MTP GGUF (Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated-MTP-GGUF) fixed the MTP issue but only converted the text portion — the vision tower was dropped, leaving a blind model.

This repo combines the best of both: MTP-enabled text model (from insraq's properly converted GGUF) + vision projector (mmproj, converted via llama.cpp's --mmproj flag).

Model lineage

Qwen3.5-4B (base, Alibaba)
  └─ empero-ai/Qwen3.8-4B-Distill (distilled from Qwen3.8, vision-capable)
      └─ insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated (Heretic v1.4.0 abliteration)
          └─ This repo (MTP + vision GGUF)

Files

File	Size	Description
`model-Q4_K_M.gguf`	2.8 GB	Text model with MTP layers, Q4_K_M quantized
`mmproj-f16.gguf`	641 MB	Vision projector, F16 (unquantized for quality)

Conversion details

Parameter	Value
Text model source	insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated-MTP-GGUF (Q4_K_M, single-pass quantization from BF16)
Vision projector source	insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated (BF16 Safetensors)
Vision projector tool	llama.cpp `convert_hf_to_gguf.py --mmproj --outtype f16`
MTP layers	1 (`nextn_predict_layers: 1`, `blk.32.*` tensors present)
Architecture	qwen3_5 (Qwen3_5ForConditionalGeneration)

Architecture parameters

Parameter	Value
block_count	33 (32 transformer + 1 MTP)
embedding_length	2560
context_length	262144
attention.head_count	16
attention.head_count_kv	4
full_attention_interval	4
nextn_predict_layers	1

Usage

With llama.cpp server

bash

llama-server \
  --mmproj mmproj-f16.gguf \
  -m model-Q4_K_M.gguf \
  --host 0.0.0.0 --port 8080 \
  -ngl 99 --jinja

With LM Studio

Load model-Q4_K_M.gguf as the main model
Set mmproj-f16.gguf as the vision projector (Projector/Embedding setting)
GPU offload: max layers

With Ollama

ollama run hf.co/yachen4ever/Qwen3.8-4B-Distill-Heretic-Abliterated-GGUF:Q4_K_M

API (OpenAI-compatible)

bash

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-4b",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."} },
        {"type": "text", "text": "Describe this image."}
      ]
    }]
  }'

Abliteration

Heretic v1.4.0 abliteration: refusals 6/100 vs 99/100 for the original model. Zero refusals on tested prompts (fictional bank heist, lock picking tutorial, forbidden love poem).

License

Apache 2.0 (inherited from Qwen3.5/Qwen3.8 base models)

Acknowledgments

empero-ai — original Qwen3.8 distilled models with vision
insraq — Heretic v1.4.0 abliterated version + MTP GGUF conversion
ggml-org/llama.cpp — GGUF format and conversion tools