The reasoner needs a vLLM server. Build a CUDA-matched env and serve Cosmos3-Nano on :8000:

just c3-doctor          # check GPU / CUDA / uv / venvs / disk + recommended CUDA pairing
just c3-setup-reason    # build the reasoner venv: vllm + vllm-cosmos3 (uv-managed)
just c3-serve-reason    # serve Cosmos3-Nano on :8000 (--max-model-len 32768)

Verify it's up before pointing an agent at it:

curl -s http://localhost:8000/v1/models    # → {"data":[{"id":"nvidia/Cosmos3-Nano",...}]}

Or one-shot a caption straight from the justfile (no Python):

just c3-reason "Caption in detail." "" scene.mp4 caption

CUDA pairing: match torch to your driver - CUDA 13 → cu130 + vllm==0.21.0; CUDA 12.8 → cu128 + vllm==0.19.1. just c3-doctor reports your driver's recommendation.

Single-GPU note: reasoner (vLLM) and generator (Diffusers) each load a 16B model and won't co-fit on one ~46GB GPU - stop one before running the other, or use separate GPUs.

uv pip install strands-cosmos                  # core: Reason2 VLM + all tools
uv pip install "strands-cosmos[cosmos3]"       # + Cosmos 3 reasoner client (vLLM server)
uv pip install "strands-cosmos[cosmos3-gen]"   # + Cosmos 3 generator (in-proc Diffusers: image/video/sound)
uv pip install "strands-cosmos[vllm]"          # + bundled vLLM + openai client
uv pip install "strands-cosmos[all]"           # everything (heavy)

The generator needs the diffusers dev build (Cosmos3OmniPipeline); PyPI forbids direct-URL deps, so pin it at install time (or just use just c3-setup-gen):

uv pip install -U "git+https://github.com/huggingface/diffusers.git"

git clone https://github.com/cagataycali/strands-cosmos && cd strands-cosmos
just setup-full    # system deps, Python deps, clones all Cosmos repos (uv-managed venvs)
just doctor        # verify everything

# dedicated, CUDA-matched envs (each is its own uv venv):
just c3-setup-reason      # reasoner: vllm + vllm-cosmos3
just c3-setup-gen         # generator: diffusers(main) + cosmos_guardrail
just c3-setup-framework   # action + training: Cosmos Framework

Run/lint/test against the dev env via uv:

uv pip install -e ".[dev]"
uv run pytest
uv run ruff check .

Drop dependencies directly into a script's header - uv run builds an ephemeral env, no manual install. Save as agent.py:

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "strands-agents[openai]",
#     "strands-cosmos",
# ]
# ///
import os, sys
from strands import Agent
from strands_cosmos import Cosmos3ReasonerModel, cosmos3_caption, cosmos3_temporal, video_probe

model = Cosmos3ReasonerModel(base_url=os.environ.get("COSMOS_BASE_URL", "http://localhost:8000/v1"))
agent = Agent(model=model, tools=[cosmos3_caption, cosmos3_temporal, video_probe])
agent(" ".join(sys.argv[1:]) or "Caption in detail: <video>scene.mp4</video>")

uv run agent.py "List the events with timestamps: <video>scene.mp4</video>"

Cosmos 3 watches a real construction-site clip, describes it, then generates new videos (one with synchronized audio) from its own description - all on a single local GPU.

① Input video

② Cosmos 3 understands it

"Two construction workers wearing yellow safety vests and helmets are walking away from the camera on a dirt path within a bustling construction site. The ground is covered in loose soil, with visible tire tracks crisscrossing the surface. In the background, a large yellow front-end loader moves slowly across the site, its bucket raised slightly as it navigates the terrain. Behind the loader, partially obscured by rebar and concrete slabs, an excavator operates near a foundation area. The scene is framed by urban buildings in the distance, including a distinctive church-like structure with a tall spire and modern glass-fronted buildings. The overall atmosphere suggests active progress on a significant infrastructure project under clear daylight conditions." Cosmos3ReasonerModel (caption in 5.2s)

The reasoner distills its own understanding into a generation prompt:

"Two construction workers in yellow safety vests and helmets walk across a dusty site, gesturing toward a yellow front loader and distant excavator as they converse."

Then Cosmos3GeneratorModel generates similar videos from that prompt (832×480, 49f):

text → video	text → video + 🔊 sound	image → video
55.5s	43.2s · AAC stereo 48kHz	42.1s · from a real frame

→ Full demo + MP4s + reasoning: demo/cosmos3_showcase/ · reproduce with uv run examples/09_cosmos3_showcase.py

just c3-setup-gen       # generator env: diffusers(main) + cosmos_guardrail
just c3-gen text2video "A robot in a warehouse." "" out.mp4

just c3-setup-framework # action + training env: Cosmos Framework
just c3-action spec.jsonl /tmp/out      # forward/inverse dynamics, policy
just c3-train-recipes                   # list SFT recipes
just c3-train vision_sft_nano           # fine-tune (8x H100); see the training guide

📖 Full guides: Cosmos 3 · Training/SFT

Dashcam safety analysis with Chain-of-Thought reasoning on Jetson AGX Thor

Performance (Jetson AGX Thor, Reason2-2B): text inference 1.4s (46 tokens) · video caption 2.2s (short clip @ 4fps), 7s load.

Jetson install:

uv pip install strands-cosmos
strands-cosmos-fix-cublas   # fix CUBLAS for Jetson GPU architecture

# Cosmos 3 Reasoner
Cosmos3ReasonerModel(base_url="http://localhost:8000/v1", reasoning=True, max_tokens=4096)

# Cosmos 3 Generator
Cosmos3GeneratorModel(model_id="nvidia/Cosmos3-Nano", guardrails=True)

# Cosmos-Reason2 VLM
CosmosVisionModel(
    model_id="nvidia/Cosmos-Reason2-8B",
    reasoning=True,           # Chain-of-thought <think>...</think>
    fps=4,
    params={"max_tokens": 4096, "temperature": 0.6},
)

strands_cosmos/
├── cosmos3_reasoner_model.py    # Cosmos3ReasonerModel (vLLM, text+vision -> text)
├── cosmos3_generator_model.py   # Cosmos3GeneratorModel (Diffusers, -> image/video/sound)
├── cosmos_vision_model.py       # CosmosVisionModel (Reason2 VLM: video+image+text)
├── cosmos_model.py              # CosmosModel (Reason2 text-only)
├── fix_cublas.py                # Jetson CUBLAS compatibility fix
├── tools/
│   ├── cosmos3.py               # 16 Cosmos 3 tools (reason/generate/action/serve)
│   ├── inference.py · reason_hf.py · serve.py        # Reason2 VLM
│   ├── predict_generate.py · transfer_generate.py    # Predict2.5 / Transfer2.5
│   ├── model_download.py · quantize.py · export_onnx.py · build_engine.py
│   ├── post_train.py · distill.py · curate.py · evaluate.py
│   └── rtp.py · nats_pub.py · video_utils.py · image_read.py · sysinfo.py
└── justfile                     # Developer workflow + c3-* recipes

CUBLAS_STATUS_INVALID_VALUE on Jetson

strands-cosmos-fix-cublas    # replaces torch's bundled CUBLAS with JetPack system CUBLAS

Cosmos 3 reasoner OOM on a single GPU - the default sequence length (262K) needs a huge KV cache. just c3-serve-reason caps it at --max-model-len 32768. Stop the generator before serving the reasoner (and vice versa).

StopIteration in get_rope_index during video (Reason2) - already handled; strands-cosmos pins a compatible transformers range. If you still see it:

uv pip install "transformers>=4.57.0,<5.3.0"

module 'torchvision.io' has no attribute 'read_video' - transformers 5.x decodes video with torchcodec and falls back to torchvision, which removed io.read_video in >=0.27:

uv pip install torchcodec

TRT tools return exit 127 - expected on workstations; those run on Jetson or in TRT Docker. Run just doctor.