Cosmos models

7 Cosmos models from NVIDIA on Hugging Face, from 2.4B to 64.6B parameters, published by NVIDIA in BF16, with contexts from 125K to 256K tokens. The smallest official model, Cosmos-Reason2-2B, needs about 5.5 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.

Pick a size

One row per official Cosmos size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.

ModelParametersNative VRAMFP8 VRAMINT4 VRAMLive GPU fit (native)Est. $/hrKV cache at 32K
Cosmos-Reason2-2B2.4B5.5 GB2.7 GB1.4 GBRTX 3070$0.088/hrnot published for this architecture
Alpamayo2-Edge2.7B5.9 GB3.0 GB1.5 GBRTX 3070$0.088/hrnot published for this architecture
Cosmos-Reason1-7B8.3B18.5 GB9.3 GB4.6 GBRTX A5000$0.176/hr1.8 GB
Cosmos-Reason2-8B8.8B19.6 GB9.8 GB4.9 GBRTX A5000$0.176/hrnot published for this architecture
Cosmos-H-Surgical15.2B33.9 GB17.0 GB8.5 GBRTX A6000$0.363/hrnot published for this architecture
Cosmos3-Super-Text2Image-4Step64.0B143 GB71.5 GB35.8 GBAMD MI300X$2.63/hr8.0 GB
Cosmos3-Super-Image2Video64.6B144 GB72.2 GB36.1 GBAMD MI300X$2.63/hr8.0 GB

VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.

Official models (7)

ModelParametersPublished asVRAM neededLive GPU fitEst. $/hr
Cosmos-Reason2-2B2.4BBF165.5 GBRTX 3070$0.088/hr
Alpamayo2-Edge2.7BBF165.9 GBRTX 3070$0.088/hr
Cosmos-Reason1-7B8.3BBF1618.5 GBRTX A5000$0.176/hr
Cosmos-Reason2-8B8.8BBF1619.6 GBRTX A5000$0.176/hr
Cosmos-H-Surgical15.2BBF1633.9 GBRTX A6000$0.363/hr
Cosmos3-Super-Text2Image-4Step64.0BBF16143 GBAMD MI300X$2.63/hr
Cosmos3-Super-Image2Video64.6BBF16144 GBAMD MI300X$2.63/hr

VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.

More from NVIDIA

Start ComfyUI on a rented GPU.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.