Particle.news

NVIDIA Expands Jetson Thor with Two Smaller Modules for Mainstream Robotics

The updates let robots run lightweight foundation models on-device, reduce memory use through automated agent skills, enable pre-release emulation for developer testing.

Overview

  • NVIDIA announced the new Jetson T3000 and T2000 modules on Wednesday, July 15, and bundled new software tools including Jetson agent skills and the Cosmos 3 Edge 4-billion-parameter robot model.
  • The T3000 is aimed at humanoids and high-end robots and combines a Blackwell GPU with an eight-core Neoverse CPU, 32 GB of LPDDR5X and 273 GB/s memory bandwidth to deliver 865 FP4 teraflops in a much smaller, lower-power package.
  • The T2000 offers a lower-entry Thor option with 400 FP4 teraflops and 16 GB of memory for visual AI agents and mobile or industrial robots that need Thor-class inference at lower cost and power.
  • Jetson agent skills automate memory and system optimizations so developers can cut memory use quickly, and NVIDIA says T3000 emulation will arrive in JetPack 7.2.1 later this month to let teams begin integration before hardware ships.
  • NVIDIA highlighted partner and early-adopter work across healthcare, logistics and manufacturing and set module shipments for general availability in Q1 2027 while noting some third-party reports have published unconfirmed specs and earlier availability claims that diverge from NVIDIA’s statements.