Paper | Website | HuggingFace

Cosmos-Reason1 is a suite of models, ontologies, and benchmarks that we develop with the goal of enabling multimodal LLMs to generate physically grounded responses. We release one multimodal LLMs: Cosmos-Reason1-7B post-trained with Physical AI SFT, and Physical AI reinforcement learning. We define ontologies for physical common sense and embodied reasoning, and also build benchmarks to evaluate Physical AI reasoning capabilities of multimodal LLMs.

Model

• Cosmos-Reason1-7B

Getting Started

Inference

from transformers import AutoProcessor
from vllm import LLM, SamplingParams
from qwen_vl_utils import process_vision_info

# You can also replace the MODEL_PATH by a safetensors folder path mentioned above
MODEL_PATH = "nvidia/Cosmos-Reason1-7B"

llm = LLM(
    model=MODEL_PATH,
    limit_mm_per_prompt={"image": 10, "video": 10},
)

sampling_params = SamplingParams(
    temperature=0.6,
    top_p=0.95,
    repetition_penalty=1.05,
    max_tokens=4096,
)

video_messages = [
    {"role": "system", "content": "You are a helpful assistant. Answer the question in the following format: <think>\nyour reasoning\n</think>\n\n<answer>\nyour answer\n</answer>."},
    {"role": "user", "content": [
            {"type": "text", "text": (
                    "Is it safe to turn right?"
                )
            },
            {
                "type": "video", 
                "video": "assets/sample.mp4",
                "fps": 4,
            }
        ]
    },
]

# Here we use video messages as a demonstration
messages = video_messages

processor = AutoProcessor.from_pretrained(MODEL_PATH)
prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
image_inputs, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True)

mm_data = {}
if image_inputs is not None:
    mm_data["image"] = image_inputs
if video_inputs is not None:
    mm_data["video"] = video_inputs

llm_inputs = {
    "prompt": prompt,
    "multi_modal_data": mm_data,

    # FPS will be returned in video_kwargs
    "mm_processor_kwargs": video_kwargs,
}

outputs = llm.generate([llm_inputs], sampling_params=sampling_params)
generated_text = outputs[0].outputs[0].text

print(generated_text)

SFT and RL Training

Please check our User Guide.

SFT and RL Training System Architecture

Cosmos-Reason1 provides toolchain to enable large scale SFT and RL training workload with following features:

1. HuggingFace Integration

   • Qwen-2.5
   • Qwen-2.5-VL
   • Qwen-3
   • Qwen-3-MoE

2. Parallelism

   • Tensor Parallelism
   • Sequence Parallelism
   • Context Parallelism
   • FSDP Parallelism
   • Pipeline Parallelism

3. Fully asynchronous (replicas specialization)

   • Policy (Consumer): Replicas of training instances
   • Rollout (Producer): Replicas of generation engines
   • Low-precision training (FP8) and rollout (FP8 & FP4) support

4. Single-Controller Architecture

   • Efficient messaging system (e.g., weight-sync, rollout, evaluate) to coordinate policy and rollout replicas
   • Dynamic NCCL Process Groups for on-the-fly replicas registration/un-registration to enable fault-tolerant and elastic large-scale RL training
   • Dynamic hyper-parameters adjustment

License and Contact

This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.

NVIDIA Cosmos source code is released under the Apache 2 License.

NVIDIA Cosmos models are released under the NVIDIA Open Model License. For a custom license, please contact [email protected].