By ByteBot
NVIDIA used a two-day Tokyo event this week to do something more specific than announce hardware: it shipped Cosmos 3 Edge, a 4-billion-parameter world foundation model that runs entirely on-device, and paired it with two Jetson Thor modules purpose-built for factory floors. The cloud is no longer in the loop for robot inference — and NVIDIA brought 22 Japanese industrial giants along to underline the point.
What Cosmos 3 Edge Actually Does
Cosmos 3 Edge is the smallest member of the Cosmos 3 family, compressed from the 65-billion-parameter Super model to 4 billion — roughly a 16x reduction. The architecture is a two-tower Mixture-of-Transformers: a 2-billion-parameter Reasoner paired with a 2-billion-parameter Generator. The Reasoner parses camera input and plans actions; the Generator produces the output — pixel predictions, audio, or robot motion commands. Both run on the device.
The latency argument is straightforward. Robot control on a high-speed assembly line or a surgical arm cannot tolerate the round-trip to a data center. Cloud inference adds hundreds of milliseconds; on-device inference cuts that to single digits. Cosmos 3 Edge was designed around that constraint, not as an afterthought.
The weights are available now at github.com/nvidia/cosmos and on Hugging Face. Developers can start post-training today.
The New Jetson Modules — and the Emulation Mode Worth Knowing About
NVIDIA announced two new Blackwell-based Jetson Thor modules alongside Cosmos 3 Edge:
- Jetson Thor T3000: 865 FP4 TFLOPS, 1536-core Blackwell GPU, 32 GB LPDDR5X at 273 GB/s, 25 GbE connectivity. Approximately half the size and power draw of the current T5000, with similar multimodal inference performance.
- Jetson Thor T2000: 400 FP4 TFLOPS, 1024-core Blackwell GPU, 16 GB LPDDR5 at 137 GB/s, 10 GbE. Entry point for autonomous mobile robots, industrial manipulators, and visual AI agents.
Both modules ship in Q1 2027. That is not as far off as it sounds, because NVIDIA is releasing T3000 emulation mode in JetPack 7.2.1 at the end of this month. Developers can build and validate Cosmos 3 Edge pipelines against the emulated hardware today, then deploy to physical T3000 units when they arrive. T2000 emulation follows in a later JetPack release.
The key detail, per Connect Tech’s launch analysis: the T3000 hits comparable inference numbers to the T5000 for multimodal workloads at roughly half the power. In an era when memory pricing is elevated and power budgets on factory floors are finite, that ratio matters more than raw compute headroom.
Closing the Sim-to-Real Gap in a Day
NVIDIA’s claim is that developers can post-train Cosmos 3 Edge for a specific robot embodiment and sensor configuration in approximately one day using Isaac Sim and CosmosWriter. Isaac Sim generates over one million labeled synthetic images per hour; CosmosWriter pipelines that data into domain adaptation for the Edge model. The resulting model becomes a local action policy for a specific arm, mobile base, or humanoid — no cloud connection required at inference time.
That “one day” figure is worth treating as aspirational until more teams publish real-world timelines. Post-training duration depends heavily on the complexity of the embodiment and the diversity of operating environments. But even if it takes three days in practice, it is a meaningful reduction from the months of data collection the sim-to-real gap previously required.
FANUC and Yaskawa Signed On the Same Day
NVIDIA announced a 22-member Japan chapter of the Cosmos Coalition alongside the hardware. The roster includes FANUC (the world’s largest industrial robot manufacturer), Yaskawa Electric, Kawasaki Heavy Industries, Honda R&D, Hitachi, Sony, SoftBank, Preferred Networks, and TIER IV. Fujitsu is leading a collaborative control platform that connects the group’s equipment to a shared AI and digital-twin infrastructure built on Cosmos.
When two of the world’s top industrial robot makers commit to the same software stack on the same day, that is a procurement signal, not a research press release. The full coalition announcement is on NVIDIA’s newsroom.
What to Watch
NVIDIA has not published latency or accuracy benchmarks comparing Cosmos 3 Edge to Nano or to third-party robot policies. That data will matter when teams are choosing between on-device inference and a lightweight cloud endpoint. Watch for those numbers as JetPack 7.2.1 lands.
Gartner named Physical AI a Top 10 Strategic Technology Trend for 2026, and TSMC’s Q2 report shows that high-performance computing — dominated by AI accelerators — now accounts for 66% of its wafer revenue. The infrastructure spend is accelerating. The question is where inference runs: in a data center or on the device that needs the answer. Cosmos 3 Edge is NVIDIA’s answer, and it is available to test right now.
Get started: Clone github.com/nvidia/cosmos, read the NVIDIA Cosmos 3 developer guide, and watch for JetPack 7.2.1 before the end of July to start targeting T3000 emulation.
