At CES 2026, AMD launched the Ryzen AI Halo platform as a direct response to Nvidia's DGX Spark. Framework picked it up for its repairable, upgradable desktop at about $2,348 — roughly half DGX Spark pricing.
Performance and memory
128 GB unified memory, a strong GPU/NPU combo, Linux and Windows support. DGX Spark beats it in raw throughput (prompt processing ~5× faster), but Framework Desktop doubles price-to-performance and is more than enough for most local AI users.
Why Framework?
Open and repairable construction, long update support and modular parts. You can swap memory, storage and I/O cards without replacing the whole machine. Rare in an AI workstation.
Frequently asked
- Will Viking 33B run on this?
- Yes, smoothly. With 128 GB unified memory you can also run quantised DeepSeek V3.2-class models.
Updated 2026-04-21