NVIDIAUnited States2025

Cosmos Reason 2

An AI model: an embodied reasoning model reasons about what a robot sees and what to do next. Its answer is typically a plan, a location or text rather than motor commands. This page describes the software, not a robot.

Robotsnot stated

VersionsCosmos-Reason2-2BCosmos-Reason2-8BCosmos-Reason2-32B

The developer has released this model in several versions. Each figure applies to the version its source names.

Key figures as stated by the developer

4 of 6 stated · 2 gaps

Model type ⁠A embodied reasoning
Controls not stated
Inputs ⁠A camera language
Access ⁠A open download
Parameters not stated
Base model ⁠A Qwen3-VL (32B size: Qwen3-VL-32B-Instruct)

How to read the figures

50HzA: The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.

The four states

Versions and notes in the key figures (6)

Model type

Developer wording: "Cosmos-Reason2: Physical AI Common Sense and Embodied Reasoning Models"; the model "can serve as a planning model to reason what steps an embodied agent might take next".

Controls

Note Model card, Output Parameters: "Text: One-dimensional (1D)"

Inputs

Model card, Input Type(s): "Text+Video/Image". Video and image = camera, text = language.

Access

Weights are downloaded from Hugging Face after accepting the licence: "By clicking Submit below, I accept the terms of the NVIDIA Open Model License Agreement".

Parameters

Note NVIDIA gives one number per size, never one for the family: the 2B and 8B sizes are announced in the repository ("The 2B and 8B models are now available on Hugging Face."), and the 32B card states "Cosmos-Reason2-32B: 32,000,000,000".

Base model

Repository: "Cosmos-Reason2 is based on the Qwen3-VL architecture." Model card, 32B size: "Network Architecture: Qwen3-VL-32B-Instruct."

Full specifications

5 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`embodied reasoning`	A Developer wording: "Cosmos-Reason2: Physical AI Common Sense and Embodied Reasoning Models"; the model "can serve as a planning model to reason what steps an embodied agent might take next".
Inputs	`camera` `language`	A Model card, Input Type(s): "Text+Video/Image". Video and image = camera, text = language.
Controls	not stated	Note Model card, Output Parameters: "Text: One-dimensional (1D)" ⁠A
Size and speed
Parameters	not stated	Note NVIDIA gives one number per size, never one for the family: the 2B and 8B sizes are announced in the repository ("The 2B and 8B models are now available on Hugging Face."), and the 32B card states "Cosmos-Reason2-32B: 32,000,000,000". ⁠A
Where it runs and what it is built on
Runs where	not stated	Note Model card: "Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems." ⁠A
Base model	Qwen3-VL (32B size: Qwen3-VL-32B-Instruct)	A Repository: "Cosmos-Reason2 is based on the Qwen3-VL architecture." Model card, 32B size: "Network Architecture: Qwen3-VL-32B-Instruct."
Openness
Access	open download	A Weights are downloaded from Hugging Face after accepting the licence: "By clicking Submit below, I accept the terms of the NVIDIA Open Model License Agreement".
License	NVIDIA Open Model License	A Model card header: "license_name: nvidia-open-model-license". "Models are commercially usable."
The developer’s own claims
Works across robots	not stated	Note The card lists robot planning as a use case: "Act as the brain for deliberate, methodical decision-making in a robot vision language action (VLA) model." ⁠A Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Not stated: Degrees of freedom controlled, Control rate, Learns new tasks

Notes

NVIDIA calls Cosmos Reason 2 an open reasoning vision language model for physical AI and robotics that can serve as a planning model for an embodied agent.
NVIDIA released the Cosmos-Reason2 models on 19 December 2025; the 32B size followed on Hugging Face on 29 April 2026.
NVIDIA now points users to Cosmos 3 and keeps this repository on limited maintenance; the Cosmos Reason 2 weights stay on Hugging Face.
NVIDIA Corporation lists its address as Santa Clara, California.

Sources

A Manufacturer product page. huggingface.co/nvidia/Cosmos-Reason2-32B · Retrieved 2026-10-14

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.