NVIDIA launches Cosmos 3 for physical AI

// 2h agoMODEL RELEASE

NVIDIA launches Cosmos 3 for physical AI

NVIDIA has released Cosmos 3, a Mixture-of-Transformers foundation platform for physical AI that consolidates visual reasoning, world generation, and action prediction into a single open-access system. Offered in 16B and 64B parameter versions, the suite includes open synthetic datasets and training tools to accelerate autonomous robotics development.

// ANALYSIS

NVIDIA is positioning itself as the standard software foundation for the physical AI era by open-sourcing a unified omni-model, effectively lowering barriers for developers while reinforcing downstream demand for its Hopper and Blackwell hardware.

–Unified Architecture: Replacing separate pipelines for perception, simulation, and planning with a Mixture-of-Transformers architecture drastically reduces complexity and latency.
–Strategic Open Access: Releasing these frontier models openly on GitHub and Hugging Face fosters a broad developer ecosystem, competing directly with closed-source robotics frameworks.
–Hardware Synergies: Optimal scaling, synthetic data generation, and enterprise fine-tuning are heavily optimized for NVIDIA's Blackwell and Hopper GPU architectures.

// TAGS

nvidianvidia-cosmos-3physical-airoboticsllmopen-sourcefoundation-modelsmixture-of-transformers

DISCOVERED

2h ago

2026-06-01

PUBLISHED

5h ago

2026-06-01

RELEVANCE

9/ 10

AUTHOR

tosh

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL33m ago

Tencent releases Hunyuan 3 Preview, a 295-billion parameter open-source Mixture-of-Experts model.

Tencent has open-sourced Hunyuan 3 Preview, a massive 295-billion parameter Mixture-of-Experts (MoE) model with an expansive 256K token context window. This model is built on an overhauled training infrastructure and introduces a hybrid fast and slow thinking approach, optimizing its performance for complex reasoning, instruction following, and advanced software engineering use cases.

UPDATE1h ago

DigitalOcean adds DeepSeek-V4-Flash to Serverless Inference

DigitalOcean has expanded its Serverless Inference platform by adding DeepSeek-V4-Flash, providing zero-infrastructure access to this 1M-token context Mixture of Experts (MoE) model. The integration features a unified OpenAI-compatible API and predictable usage-based pricing to easily scale high-throughput, agent-based workloads.

SECURITY1h ago

A supply chain attack dubbed Miasma has compromised over 30 official Red Hat npm packages with credential-stealing malware.

A highly sophisticated supply chain attack hijacked a Red Hat developer account to inject the Miasma malware into more than 30 official npm packages under the @redhat-cloud-services scope. During installation, these backdoored packages silently harvest sensitive information, including cloud logins, CI/CD secrets, and Kubernetes tokens, presenting a severe threat to organizations using these services.