GitHub
Menu

Models / NVIDIA

Cosmos3-Nano
serving recipe.

Cosmos3-Nano is a 16B dense model with a 256K tokens context window. vLLM publishes a tested serving recipe for it, and Wingsmith reads that recipe when you register the weights.

Size
16BDense
Context
256K tokensAs the recipe states it.
Memory floor
39 GBSmallest published, whole replica.
vLLM
0.21.0 or laterMinimum version in the recipe.

01 / Memory

Precisions and memory floors.

Each variant is a precision vLLM tested, sometimes published as its own repository.

VariantPrecisionMemory floorWeights
defaultBF1639 GBnvidia/Cosmos3-Nano

Floors are the minimum the vLLM recipe publishes across the whole replica, from vLLM's own runs. The control plane compares them with its own sizing from the model config.

02 / Parallelism

Layouts the recipe supports.

The recipe does not list parallelism strategies; its flags carry the tested layout.

03 / Hardware

Where vLLM ran it.

Node shapes the recipe reports an end-to-end run on.

  • H100 · 8 GPUs
  • H200 · 8 GPUs

Wingsmith runs models through NVIDIA Dynamo, so it serves on NVIDIA GPUs. Other vendors are listed as the recipe states them.

04 / Flags

The tested baseline.

Base flags, then the default variant, then the features that are on by default, as the recipe composes them.

vllm serve nvidia/Cosmos3-Nano \
  --host 0.0.0.0 \
  --port 8000 \
  --init-timeout 1800
Single-node baseline. Multi-node layouts add flags per strategy and role.

05 / Serve

Serve it with Wingsmith.

From a slug to an OpenAI-compatible endpoint, across your clusters.

  1. Register the weights

    In the console, or through the API. Resolving a slug proposes an artifact and writes nothing.

    curl "http://localhost:8000/v1/artifacts/resolve?source=hf:nvidia/Cosmos3-Nano"
    Against the local control plane from the quickstart, where admin auth is off.

  2. Read the draft

    The draft carries this recipe as upstream evidence: the tested flags, the precision and memory floor, the parallelism layouts, and the hardware vLLM verified, with links back to the recipe. Nothing from it is applied without you, and the draft warns when the plane's own sizing and the published floor differ by a quarter or more.

  3. Create the workload

    Choose the engine, device shape and scale. The plane picks the clusters, and each cluster's operator renders the deployment as a Dynamo graph.

  4. Give it an endpoint

    One model name for clients, routed by the gateway to deployments on every cluster serving it.

Run the control plane locally →

Questions

Common questions.

  • How much GPU memory does Cosmos3-Nano need?The vLLM recipe publishes a minimum of 39 GB across the whole replica for its BF16 variant. That is vLLM's own measurement, not a per-GPU figure.
  • Which parallelism layouts does the Cosmos3-Nano recipe support?The recipe does not list parallelism strategies.
  • Can Wingsmith serve Cosmos3-Nano?Wingsmith serves models through NVIDIA Dynamo on clusters with NVIDIA GPUs. Register nvidia/Cosmos3-Nano and the plane attaches this vLLM recipe to the draft as evidence: its tested flags, precision and memory floor, for you to review rather than applied on your behalf.

From the vLLM recipe for nvidia/Cosmos3-Nano ↗ (snapshot at 4bcc189), Apache 2.0.

Every model

Status

Open source,
under Apache 2.0.

Wingsmith is pre-release. The first public release, and everything after it, will be under Apache 2.0.

View on GitHubRead the docs Run it locally

Built and not built

Built

Multi-cluster placement and routing, Dynamo deployments, device inventory, endpoints and quotas, the console, the CLI, and three measurement tools.

Not built

Published images and charts, per-user roles and SSO, batch and training workloads, and a Helm chart for the control plane.

Pre-release. No tagged release, images, or charts yet, and APIs and protocols can change.