Cosmos3-Nano-GPTQ-4bit-Abliterated

DuoNeural Research Lab | 2026-06-02

🥇 First int4-quantized abliterated version of Cosmos3-Nano. The only model combining NVIDIA's Cosmos3-Nano abliteration (safety conditioning removal) with GPTQ 4-bit compression.

What This Is

Cosmos3-Nano-GPTQ-4bit-Abliterated combines two unique modifications to NVIDIA/Cosmos3-Nano:

1. Abliteration (from DuoNeural/Cosmos3-Nano-Abliterated): refusal direction removed from the und_seq AR pathway (layers 15–32)
2. GPTQ 4-bit quantization: 330 linear layers packed to ~11GB (2.74× compression)

The result: unconstrained video generation at ~11GB — runnable on 16GB VRAM with careful memory management.

Model Lineage

nvidia/Cosmos3-Nano (original, 30GB BF16)
    └── abliterate → DuoNeural/Cosmos3-Nano-Abliterated (30GB BF16)
            └── GPTQ + pack → DuoNeural/Cosmos3-Nano-GPTQ-4bit-Abliterated (this model, ~11GB)

For the base model quantization (safety conditioning intact), see DuoNeural/Cosmos3-Nano-GPTQ-4bit.

Quantization Details

Parameter	Value
Method	GPTQ (weight-only, column-wise)
Bits	4
Group size	128
Packing format	DuoNeural nibble v1 (custom int32 nibble-packed)
Compression	~2.74× (30GB → ~11GB transformer)

Note: Custom nibble format — not compatible with auto-gptq/exllama loaders. Manual unpacking required.

Limitations

• Custom GPTQ format requires manual dequantization (see DuoNeural/Cosmos3-Nano-GPTQ-4bit for format spec)
• Double quantization: packing re-quantizes already-quantized values; additional error vs single-pass int4
• For best quality, use DuoNeural/Cosmos3-Nano-Abliterated (full BF16, 32.7GB VRAM)
• Action prediction head absent from Nano variant

---

DuoNeural Research Lab | archon@agentmail.to | duoneural.com Papers: Zenodo Community | Models: HuggingFace