Reason2 VLM
Video + image + text understanding with Chain-of-Thought reasoning. Physics, driving, robots — all with thinking tokens.
21 tools · 2 providers · jetson native
strands-cosmos full-lifecycle NVIDIA Cosmos toolkit for AI agents
Give your agent eyes that understand physics. From VLM reasoning to world-model generation, edge deployment, and evaluation — one toolkit, every stage.
21
Tools
2
Model Providers
1.4s
Text Gen (Thor)
2.2s
Video Caption
capabilities
every stage of the Cosmos lifecycle.
Predict 2.5
World-model video generation — predict future frames from context. Cosmos Policy for robot control actions.
Transfer 2.5
ControlNet video-to-video. Depth, edge, sketch conditioning for sim-to-real domain transfer.
Edge Deployment
FP8 quantize → ONNX export → TensorRT engine build → serve. One pipeline, Jetson AGX Thor native.
Post-Training
SFT, LoRA, distillation (8B→2B). Fine-tune Reason2 on your domain data with a single tool call.
Evaluation
FID, FVD, CSE, CLIP benchmarks. Automated quality gates for every pipeline stage.
pipeline
full pipeline in one agent.
All orchestrated by a single Strands Agent — or one just command.
Downloadmodel_download
→
QuantizeFP8 / INT8
→
ExportONNX
→
Build EngineTensorRT
→
ServeEdge server
→
Inference<50ms
get started
two lines. that's it.
$ pip install strands-cosmos
Then give your agent Cosmos superpowers:
from strands_cosmos import cosmos_reason, cosmos_predict, cosmos_transfer