NVIDIA

AI Hardware · Laptops · Local AI Models

Launch

NVIDIA RTX Spark laptops run a 74B-parameter AI model in 24GB

Sparse mixture-of-experts routing and quantization-aware training squeeze frontier-scale compute onto hardware that never touches the cloud.

• NVIDIA detailed the full software stack for its RTX Spark laptop platform, including the Nemotron 3.5 Bolt 74B model that releases Oct. 15, one day ahead of the laptops going on sale Oct. 16.
• RTX Spark is NVIDIA's first unified 'superchip' laptop platform, pairing a Grace CPU with a Blackwell GPU to deliver up to 128GB of shared memory and 1 petaflop of AI compute for gaming, creative work and local AI agents.
• Nemotron 3.5 Bolt 74B is a mixture-of-experts model that activates only about 9 billion of its 74 billion parameters per token, then compresses to 24GB using 2-bit quantization-aware training rather than standard post-training compression.
• The platform's NVLink-C2C interconnect moves data between CPU and GPU at 900 GB/s, versus 64 GB/s for a PCIe Gen 5 x16 slot — the bandwidth that makes a single pool of unified memory usable for large models.
• NVIDIA's PAIR router exits beta and can auto-discover other GPUs on a home network to split inference across machines, while a sandboxed local-agent execution framework built with Microsoft enters early preview.
• NVIDIA says Adobe has rearchitected Photoshop and Premiere for up to 2x performance on RTX Spark, and hundreds of games with DLSS 5 and major anti-cheat support are ready at launch, positioning the device as more than an AI niche product.
• By open-sourcing a frontier-scale model built to run entirely offline, NVIDIA is betting that local inference — not cloud subscriptions — becomes the default way consumers access high-end AI agents.