rf-detr vs cosmos
Two of the top projects in Open models, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
rf-detr
Real-time transformer models for object detection, segmentation and keypoints
cosmos
Open world foundation models for robotics and driving: video, reasoning, policies
Verdict
rf-detr ranks higher, 85 to 72. Of the 4 score parts we measure for both, rf-detr leads on 2, cosmos on 1 and 1 is even.
- Adoptioncosmos leads by 6 points, 73 to 67.
- FreshnessEven at 100.
- Maintenancerf-detr leads by 36 points, 93 to 57: a wide gap.
- Agent-readyrf-detr leads by 40 points, 70 to 30: a wide gap.
Pick rf-detr if you are
Computer vision engineers needing real-time detection or segmentation models
- Detection, segmentation and keypoints share one Python API across Nano to 2XLarge sizes
- Apache 2.0 for the package and N/S/M/L detection and all segmentation models
Pick cosmos if you are
Robotics and autonomous-vehicle teams building on open world models
- Three sizes: 64B Super, 16B Nano, 4B Edge for Jetson and RTX Pro 6000
- Runs on Diffusers, Transformers, vLLM, vLLM-Omni, SGLang, TensorRT-LLM and NIM
| What we compare | rf-detr | cosmos |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 67, popular | 73, popular |
| Freshness | 100, active | 100, active |
| Maintenance | 93, healthy | 57, fair |
| Agent-ready | 70, partly | 30, minimal |
| Facts from GitHub and the README | ||
| Stars | 10.1k | 12.4k |
| License | Apache-2.0 (permissive) | custom license (read the license) |
| Last commit | Nov 2026 | Nov 2026 |
| Last release | Nov 2026 | Jun 2026 |
| Language | Python | Jupyter Notebook |
| Docker | No | No |
| GPU | Optional | Required |
| arm64 or Apple Silicon | Not stated | Not stated |
rf-detr
RF-DETR is a Python package (`pip install rfdetr`, Python 3.10+) from Roboflow that provides pretrained DINOv2-based transformer models for object detection, instance segmentation and keypoint detection (preview). Models come in sizes from Nano to 2XLarge, run through one API via classes like `RFDETRMedium`, and can be fine-tuned on custom datasets. Some sizes (Atto, Femto, Pico, XL, 2XL) are under the PML 1.0 license rather than Apache 2.0.
Who it is for: Computer vision engineers needing real-time detection or segmentation models
Strengths
- Detection, segmentation and keypoints share one Python API across Nano to 2XLarge sizes
- Apache 2.0 for the package and N/S/M/L detection and all segmentation models
- README benchmarks on COCO and RF100-VL, with T4 TensorRT FP16 latency per size
- Fine-tuning supported, with Colab notebook and Hugging Face Space provided
Weaknesses
- Atto, Femto, Pico, XL and 2XL detection models use PML 1.0, not Apache 2.0
- Keypoint detection is labeled preview
- Latency figures are T4 TensorRT only; CPU and edge numbers not given in the visible README
- NAS for your own dataset requires the Roboflow platform
- GPU optional
- Needs PyTorch, supervision
- Models: RF-DETR detection (Atto to 2XL), RF-DETR instance segmentation (Seg-N to Seg-2XL), RF-DETR keypoint (preview), DINOv2 backbone, COCO and custom datasets
cosmos
Landing repo for Cosmos 3, a family of omnimodal world models with a Reasoner (text and vision to text) and a Generator (text, vision, audio, action to vision, audio, action). Ships three base sizes (Super 64B, Nano 16B, Edge 4B) plus post-trained checkpoints for driving simulation, transfer and DROID robot policies. The repo holds a cookbook of notebooks; training, curation and evaluation live in separate Cosmos repos.
Who it is for: Robotics and autonomous-vehicle teams building on open world models
Strengths
- Three sizes: 64B Super, 16B Nano, 4B Edge for Jetson and RTX Pro 6000
- Runs on Diffusers, Transformers, vLLM, vLLM-Omni, SGLang, TensorRT-LLM and NIM
- Post-trained checkpoints for driving simulation, multi-view transfer and DROID robot policy
- Companion repos cover SFT, LoRA, RL, distillation, data curation and evaluation
Weaknesses
- Target hardware is H100, H200, B200, Jetson or RTX Pro 6000 class GPUs
- Quickstart needs gated access to the Cosmos-1.0-Guardrail repo on Hugging Face
- License is OpenMDW-1.1, not a common OSI license; custom licensing by email
- README warns of artifacts in long, high-resolution or physically complex outputs
- GPU required
- Needs Hugging Face account and token, Cosmos-1.0-Guardrail (gated), PyTorch, diffusers
- Models: Cosmos3-Super 64B, Cosmos3-Nano 16B, Cosmos3-Edge 4B, Diffusers, Transformers
More in Open models
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.