The scorecard for the Agentic AI PC category.
RTXsparks.com is the independent registry and scorecard for NVIDIA RTX Spark AI PCs — helping you decide what to buy, which agents to run, and how far you can scale. We benchmark every shipping Grace-Blackwell system for tokens/sec, agent stability, and lifecycle trust.
Buy
Independent pricing and scoring for every shipping Spark laptop, mini-PC, and desktop. Filter by your workload.
Explore →02Run
Opinionated model menus from 7B to 235B and curated agent stacks calibrated for unified memory.
Explore →03Scale
Multi-node Spark cluster playbooks. From a 2×64GB desk pair to 4-node racks.
Explore →Top Performers
Ranked by Overall Score (v2026.4.1)
Find your break-even
At 50k tokens/day on Llama-3.1-70B, a 128GB Spark workstation pays back versus an H100 cloud instance in roughly 6 months.
Model menu
Gold standard multi-agent reasoning. 18–24 tok/s on Grace-Blackwell.
Top open coder. 34–41 tok/s, 32k context.
Reference NVIDIA-aligned agent stack with NPU offload.
We buy our own hardware, publish our scoring weights, version our benchmarks, and accept no OEM money. Read the methodology.