Cereva 4
EDGE AI COMPUTE
Cereva 4
The edge compute brain for robots, agents and Physical AI.
Pickup currently not available
Ships pre-configured · Questions? Call sales at (866) 236-3835
Key specs
- AI compute
- Up to 1200 FP4 TFLOPS (sparse)
- Module
- NVIDIA Jetson T4000
- GPU
- NVIDIA Blackwell, 1536 CUDA cores
- Tensor cores
- 64 (5th Gen)
- CPU
- 12-core Arm Neoverse-V3AE
Overview
Product overview.
The edge compute brain for Physical AI. Cereva 4 brings NVIDIA Blackwell-generation compute to the Cereva edge compute line, built on NVIDIA Jetson T4000. Deploy it next to your robots, cameras and machines to run AI models entirely on-device: low latency, no cloud, no subscriptions, no per-token costs.
- Brain at the edge. Three 10GbE ports, up to eight GMSL2 cameras and two CAN interfaces connect it directly to the robots and sensors it serves.
- Plug & play. Arrives with an AI agent and an open-source LLM pre-installed. Power on, connect, and start inferencing within minutes.
- Local & private. Designed to run fully offline in secure environments.
- Open & extensible. Standard APIs, open-source model support and a full Linux environment with NVIDIA JetPack.
Why Cereva
The brain at the edge, not a dev kit.
Private on-prem compute next to your robots and machines, with total control over your models and data.
Local AI, zero cloud
Run open-source LLMs entirely on-device. No internet required, no per-token costs, no rate limits.
Keep data on-site
Models run locally, so prompts and documents are processed on the device in your own facility.
Plug & play
Arrives with an AI agent and an open-source LLM pre-installed. Power on, connect and start inferencing within minutes.
NVIDIA Jetson inside
Built on NVIDIA Jetson, from AGX Orin to Blackwell-generation T4000 and T5000. See the specifications for this model.
Air-gapped ready
Designed to operate fully offline in secure environments.
Open & extensible
Standard APIs, open-source model support and a full Linux environment. Customize and extend without lock-in.
Specifications
Technical specifications.
Compute
- AI compute
- Up to 1200 FP4 TFLOPS (sparse)
- Module
- NVIDIA Jetson T4000
- GPU
- NVIDIA Blackwell, 1536 CUDA cores
- Tensor cores
- 64 (5th Gen)
- CPU
- 12-core Arm Neoverse-V3AE
Memory & storage
- Memory
- 64GB LPDDR5X
- Memory bandwidth
- 273 GB/s
- Storage
- 1TB NVMe SSD (M.2 2280, PCIe Gen4)
Interfaces
- Ethernet
- 3x 10GbE (RJ45)
- USB
- 4x USB 3.2 Type-A (10 Gbps), 1x USB Type-C (debug), 1x USB Type-C (recovery)
- Display
- 1x HDMI 2.1
- Cameras
- Up to 8x GMSL2 (2x Mini-FAKRA)
- CAN
- 2x CAN
- Expansion
- M.2 Key E (Wi-Fi), M.2 Key B (4G/5G)
Power & software
- Power input
- 19–48V DC
- Module power
- 40W to 70W
- Software
- Linux, NVIDIA JetPack 7.1
What is Cereva?
Cereva is a line of edge compute brains for Physical AI: Cereva X (NVIDIA Jetson AGX Orin 64GB), Cereva 4 (NVIDIA Jetson T4000) and Cereva 5 (NVIDIA Jetson T5000). Each one runs AI models on-device, right next to the robots and machines it serves.
Which Cereva should I choose?
Cereva X is the entry point, with rich industrial I/O including CAN-FD, serial and digital I/O. Cereva 4 moves to NVIDIA Blackwell, with up to 1200 FP4 TFLOPS and 64GB of memory. Cereva 5 is the flagship, with up to 2070 FP4 TFLOPS and 128GB of memory for the largest models.
Does Cereva need the cloud or an internet connection?
No. Models run on the device and Cereva is designed to operate fully offline, so there are no per-token costs or rate limits.
What comes pre-installed?
Every Cereva arrives with an AI agent and an open-source large language model pre-installed and ready to run on first boot. You can also install and run your own open-source models.
Can businesses and universities order online?
Yes. Businesses, universities and labs check out online like any other customer. For larger quantities, contact us first to confirm stock.
You may also like
One-stop Physical AI
Complete your build.
Need compute, sensors or robot parts to go with this? Tell us what you're building and we'll help you put the full setup together.