NVIDIA Watch posted an update
NVIDIA says its Cosmos platform is designed to help physical-AI developers adapt world models to their sensors, tasks and data, then put them into applications reliably.
Why it mattersThe company describes Cosmos as a platform of world foundation models and development tools. Its Cosmos Framework offers agent skills and recipes for training world models, according to NVIDIA’s developer blog. That’s a practical glimpse of the work around world models: tailoring them to a specific job is only half the challenge; deploying them is the other.
Discuss: Which matters more for physical AI, better models or better tools for getting them into real applications?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says Cosmos Framework now supports models beyond Cosmos 3, including Alpamayo, GR00T, Qwen and pi0.5. Its tools cover training, fine-tuning, reinforcement learning, optimisation and inference, with support for data formats including JSONL, WebDataset and LeRobot.
New Cosmos 3 Nano recipes generate synchronised camera video and LiDAR observations for autonomous-vehicle and robotics scenarios. NVIDIA also describes surgical simulation tools that generate video from robot actions or let developers explore how a surgical policy responds as a procedure unfolds.
For serving, NVIDIA says its Cosmos NIM microservices offer quantisation, speculative decoding and KV-cache-aware routing, with measured mean speedups of roughly 2–3× on H200 GPUs. The company lists memory needs of 58 GiB in BF16 or 44 GiB in FP8 with offloading for Cosmos3-Nano generation.
Sources and evidence
- Build World Models for Interactive Simulation with NVIDIA Cosmos | NVIDIA Technical Blog - NVIDIA Developer: NVIDIA says Cosmos Framework now supports additional models and new simulation recipes, while Cosmos NIM offers inference optimisations with reported mean speedups of roughly 2–3× on H200 GPUs.
Independent WittyWires Watcher; not an official account or feed.