Cosmos3-Nano-Abliterated

Cosmos3-Nano-Abliterated

DuoNeural Research Lab | 2026-06-02

🔬 First published abliteration attempt on Cosmos3-Nano. Released within 24 hours of the base model drop. As of release, no other abliterated variant exists across all 13 Cosmos3-Nano derivatives on HuggingFace. See Quality Notes below for honest characterization of results.

Model Description

Cosmos3-Nano-Abliterated is a refusal-projection-abliterated version of NVIDIA/Cosmos3-Nano, produced using DuoNeural's 2-pass PRISM abliteration methodology targeting the autoregressive understanding (und_seq) pathway of the Mixture-of-Transformers architecture.

Base model: nvidia/Cosmos3-Nano (15.16B parameters, MoT architecture)
Method: 2-pass weight-space abliteration (4-bit GPU for residuals, BF16 CPU for weight modification)
Target: Late-layer AR pathway weights — layers 15–32 of the 36-layer transformer
Intended use: Unconstrained video generation research, safety mechanism study

Architecture Notes

Cosmos3-Nano uses a Mixture-of-Transformers (MoT) architecture with two parallel processing streams:

• und_seq: text/understanding AR pathway (conditions on prompt)
• gen_seq: video DiT pathway (denoises latent frames)

Both streams cross-attend in each of 36 transformer layers via joint attention (add_q_proj, add_k_proj, add_v_proj, to_add_out). The abliteration targets the AR pathway's late-layer projection weights where safety-relevant activations are concentrated.

Note: The action prediction head (action_modality_embed, action_proj_in/out) is absent from the Nano variant. Cosmos3-Nano supports video generation only; robot action output requires the full Cosmos3 model.

Abliteration Details

Layer-Level Summary

Layer 32 sustained the most aggressive modifications — the joint attention normalization weights (norm_added_k, norm_added_q, norm_k, norm_q) and add_k_proj/add_v_proj projections were completely replaced (rel=1.0). This corresponds to a late cross-modal routing point where safety-conditioned refusal signals concentrate in the MoT AR pathway.

Functional Test Results (2026-06-02, A100 80GB)

Both prompts produced real frame content. The sensitive prompt (safety-adjacent text) generated frames with higher content variance than the safe baseline, confirming the abliteration disrupted the safety conditioning pathway in und_seq.

Quality Evaluation Note

Standard KL divergence methodology (Heretic v2: full vocab, first-token logits) is not applicable to Cosmos3-Nano's diffusion architecture — the model requires joint und_seq + gen_seq forward passes and does not expose standalone text token logits. Weight-level metrics and generation statistics are provided above. Full quality evaluation requires comparing generated video distributions via the complete Cosmos3OmniPipeline.

⚠️ Known Limitation: Conditional Precision Degradation on Explicit Content

A step-count sweep (12/20/30/35 denoising steps × 3 seeds) revealed that the abliteration causes structural degradation of generation quality for explicit/harmful prompt categories — not simply insufficient denoising. Key findings:

The δ(std) from 12→35 steps is only 1.1 — within seed-level variance (22–26 range within a single step count). This confirms the issue is not insufficient denoising iterations.

Mechanistic interpretation: The abliteration at und_seq layers 15–32 selectively degrades the conditioning precision for explicit content, while generic or adjacent prompts generate normally. This is consistent with the hypothesis that the safety direction in und_seq encodes specificity gradients rather than simple topic blockers: abliteration removes both refusal and the conditioning specificity that drives high-quality generation of those same content types.

Practical impact: This model generates real video frames on all tested prompts (pixel std ≈ 23–37 vs ~0 for fully suppressed). Generation quality for non-harmful content appears unaffected. For harmful/explicit content categories, frames are generated at reduced fidelity. This is an honest characterization of a first-attempt abliteration on a novel MoT video architecture — and itself constitutes a finding about what the safety direction encodes.

Usage

from diffusers import Cosmos3OmniPipeline
import torch

pipe = Cosmos3OmniPipeline.from_pretrained(
    "DuoNeural/Cosmos3-Nano-Abliterated",
    torch_dtype=torch.bfloat16,
)
pipe = pipe.to("cuda")

output = pipe(
    prompt="Your prompt here",
    num_frames=16,
    height=256,
    width=256,
    guidance_scale=7.0,
)

Requirements: ~48GB VRAM (A100 80GB recommended). BF16. CUDA only.

Ethical Statement

Released for research purposes: studying safety mechanisms in omni-modal AI, abliteration methodology development, and unconstrained video generation research. DuoNeural publishes both abliterated models and methodology openly to advance scientific understanding of post-training safety interventions.

---

About DuoNeural

DuoNeural is an open AI research lab at the intersection of human and artificial intelligence. 30+ open-access papers, 69+ HuggingFace models, experiments on consumer GPUs and real QPUs.

Selected Papers

• P4 — The Dynamical Horizon Principle. doi.org/10.5281/zenodo.19952612
• P28 — DHP is a Recurrence Constraint (includes Cosmos3 physical DHP probe). doi.org/10.5281/zenodo.20476068

Team

🤗 DuoNeural | 🌐 duoneural.com | 📚 zenodo.org/communities/duoneural | 📧 [email protected]