cosmos-reason2-2b-text-only

Cosmos-Reason2-2B Text-Only

This is a text-only extraction of nvidia/Cosmos-Reason2-2B.

Built on NVIDIA Cosmos.

The original model is a Qwen3-VL/Cosmos vision-language model. This repository keeps the language backbone and lm_head, removes the vision tower and projector weights, and saves the result as a standalone Hugging Face Qwen3ForCausalLM checkpoint.

What Changed

• Source model: nvidia/Cosmos-Reason2-2B
• Output architecture: Qwen3ForCausalLM
• Output model_type: qwen3
• Kept tensors: 311
• Dropped tensors: 315
• Removed weight prefixes include model.visual.* and other multimodal components
• Output weights: model.safetensors

The original nested text config used qwen3_vl_text. It was converted to a qwen3 CausalLM-compatible config because the tested Transformers environment did not expose qwen3_vl_text through AutoModelForCausalLM.

Validation

Validated locally with:

• torch 2.12.1+cpu
• transformers 5.12.1
• safetensors 0.8.0

Checks performed:

• AutoConfig.from_pretrained(...) loads as Qwen3Config
• AutoTokenizer.from_pretrained(...) loads as Qwen2Tokenizer
• AutoModelForCausalLM.from_pretrained(...) loads as Qwen3ForCausalLM
• Forward pass succeeds on a short text prompt
• Output logits shape: (1, 7, 151936)
• No visual, vision, projector, or language_model tensor names remain in the exported checkpoint

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sasa2000/cosmos-reason2-2b-text-only"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

inputs = tokenizer("Explain why objects fall toward Earth.", return_tensors="pt").to(model.device)
with torch.no_grad():
    output_ids = model.generate(**inputs, max_new_tokens=128)

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Limitations

This checkpoint is text-only. It does not include the original vision tower, video/image processor, or multimodal projector weights. Image and video inputs are not supported.

This is an unofficial derived checkpoint and is not released by NVIDIA.

License

The source model is released under the NVIDIA Open Model License. Use of this derived checkpoint must comply with the original model license and any applicable terms.

Licensed by NVIDIA Corporation under the NVIDIA Open Model License.