NVIDIA RTX Spark: 128GB Unified Memory for On-Device AI
NVIDIA has announced RTX Spark, a mini-superchip designed for laptops and small desktops: Blackwell GPU with approximately 6,100 cores, up to a 20-core CPU, around 1 PFLOP FP4 compute, and up to 128GB unified memory. The 128GB configuration is enough to run models like Qwen 3.6 35B (~72GB) fully locally. RTX Spark ships in fall 2026.
Why It Matters
128GB of unified memory at laptop form-factor is the hardware enabler that makes running frontier-class open models like Nemotron 3 Ultra's active layers fully on-device a consumer-grade proposition — not just a data center one — by end of 2026.