How to use from the
Use from the
Diffusers library
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("mwnuk/Cosmos3-Edge-severed", dtype=torch.bfloat16, device_map="cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

Cosmos3-Edge — severed towers (self-contained)

The two streams of the nvidia/Cosmos3-Edge Mixture-of-Transformers, extracted as independently loadable components, PLUS unmodified copies of every supporting part needed to recreate the pipeline end-to-end (no dependency on the original repo). Weights are unmodified slices/copies of the original checkpoint (original tensor names preserved; see manifest.json in each component). Derivative of NVIDIA Cosmos3-Edge under the OpenMDW-1.1 license.

Contents

Folder Contents Size
denoiser/ Extracted generation stream: add_* attention, *_moe_gen MLPs/norms, latent I/O (proj_in/out), timestep embedder, domain-aware action heads 2.85 GB
reasoner/ Extracted understanding stream (VLM): embed_tokens, lm_head, und attention/MLPs/norms 3.89 GB
vae/ Wan VAE (video ↔ 48-ch latents, 4× temporal / 8× spatial) — required at the denoiser boundary 1.4 GB
vision_encoder/ SigLIP2 + projector — only needed for image/video understanding prompts to the reasoner 1.0 GB
text_tokenizer/ Tokenizer + chat template — reasoner input ~20 MB
scheduler/, root configs UniPC scheduler config, model_index.json, transformer/processor configs tiny
assets/ The pinned example inputs (i2v image + JSON prompts) used by the verification scripts ~1 MB

Both component folders include the 28 per-layer k_norm_und_for_gen interface norms (the coupling surface between the towers), buffers.safetensors (non-persistent rotary inv_freq, fp32), and a copy of the original transformer config.

Not included: the original fused transformer/ shards (6.7 GB) — by construction the two component files reproduce it exactly; fetch nvidia/Cosmos3-Edge if the fused model is needed. model.safetensors.index.json (the transformers-side index) is kept for reference and refers to those original shards.

Verification (bit-exactness)

  • Cosmos3's MoT attention is one-way: und tokens self-attend causally; gen tokens attend over concat(und_KV, gen_KV). The und-KV for a fixed prompt is constant across denoising steps (verified bit-identical, 28/28 layers).
  • Denoiser rebuilt from denoiser/ alone + cached und-KV reproduces the full 4B model's output bit-for-bit (max abs diff 0 on decoded video).
  • Reasoner rebuilt from reasoner/ alone reproduces the full model's per-layer und-KV bit-for-bit (28/28 layers), from bare input_ids.
  • Two build requirements for exactness: run time_embedder in float32 (_keep_in_fp32_modules), and load buffers.safetensors (a bf16-instantiated rotary inv_freq diverges).

Usage

Loader/verification scripts and the full architecture analysis live in github.com/wnukmat/cosmos (a private repository) (scripts/standalone_severed.py, scripts/reasoner_kv.py, EXTRACTION_PLAN.md). Requires diffusers >= 0.41.0.dev0 (Cosmos3 classes). Purpose: reference implementation to test a 3-tower world-action-model rebuild (external VLM + this denoiser + custom action expert) against known-good components.

License

The weights in this repository are unmodified slices and copies of nvidia/Cosmos3-Edge, which NVIDIA releases under the OpenMDW License Agreement, version 1.1. This repository is distributed under the same agreement; a copy is in LICENSE. Notices of origin from the original model apply. This is not an NVIDIA release and is not endorsed by NVIDIA.

Downloads last month
24
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for mwnuk/Cosmos3-Edge-severed

Finetuned
(28)
this model