The first RTX Spark computer goes on sale, capable of running large models with 120 billion parameters.

36Kr reports that the first NVIDIA RTX Spark computer has gone on sale and highlights its ability to run large models with up to 120 billion parameters. The claim reflects the platform's large unified-memory design, which combines an Arm-based Grace CPU and Blackwell GPU resources in a system built for local AI and other memory-intensive workloads.

Model capacity is not the same as model speed or overall usability. Actual results depend on quantisation, context length, memory bandwidth, software frameworks, CUDA support, cooling, power limits, and whether the model and tooling are optimised for the platform. Buyers should look for reproducible benchmarks and verify the exact configuration before treating a 120-billion-parameter claim as a guarantee for every workload.

For Canadian developers, the launch also leaves practical questions about local pricing, availability, taxes, warranty support, shipping, and software access. Manufacturer listings and independent testing should be checked before comparing an RTX Spark system with a cloud GPU, an x86 workstation, or Apple silicon. Running models locally may improve privacy or reduce network dependence, but it does not automatically mean lower cost or faster responses.