Cosmos3-Super
Support summary
Cosmos3-Super is a checkpoint published by NVIDIA, built on the Cosmos3ForConditionalGeneration architecture. vLLM 0.28.0, the newest indexed vLLM release, documents support.
vLLM documents support in 4 of 20 indexed releases, from 0.26.0 to 0.28.0.
- Documented hardware
- NVIDIA / CUDA
- Documented quantization
- BF16
- Documented features
- Tensor parallelism
- Configured context
- 262,144 tokens
- License
- openmdw1.1-license
Hardware, quantization and features count only statements about this checkpoint; engine-wide capabilities are listed with each release.
Hugging Face model card
- Publisher
- NVIDIA
- License
- openmdw1.1-license
- Library
- cosmos
- Architecture
- Cosmos3ForConditionalGeneration
- Weight tensor elements
- 64.6B
- Tensor types
- BF16
- Configured context
- 262,144 tokens
- Weight access
- Public
- Repository created
- 2026-03-10
Model notes and sources
- vLLM v0.28.0 · docs/models/supported_models.md · 4 indexed revisions, v0.26.0 to v0.28.0
- nvidia/Cosmos3-Super · publisher model card
- nvidia/Cosmos3-Super · Hugging Face metadata
- nvidia/Cosmos3-Super · publisher configuration
- vllm · models/nvidia/Cosmos3-Super.yaml
- sglang · docs/cookbook/diffusion/Cosmos/Cosmos3.mdx
- Reviewed
Engine support
3 engines
| Engine | Release | Checkpoint support | Evidence scope | Last known supported |
|---|---|---|---|---|
| vLLM | 0.28.0Newest indexed | Checkpoint evidenceSource: vLLM v0.28.0 · docs/models/supported_models.mdReviewed | 0.28.0This release | |
Features, dependencies and evidence for vLLM 0.28.0 | ||||
| SGLang | 0.5.19Newest indexed | No matching evidence curated | No supported release indexed | |
Features, dependencies and evidence for SGLang 0.5.19 | ||||
| TensorRT-LLM | 1.2.1Newest indexed | No matching evidence curated | No supported release indexed | |
Features, dependencies and evidence for TensorRT-LLM 1.2.1 | ||||