Revision History
NVIDIA DGX Spark · 14 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Update: added an official video.
Version 14 · Nov 1, 2026, 08:02 AM
- Update: NVIDIA FP4-Local checkpoints (GLM-5.3-Flash, Qwen3.8-Flash-Next) validated on DGX Spark; $6,950 128 GB price (Oct 2, 2026) per The Register/Techno-Edge; new M5 Ultra Mac Studio comparison with memory-bound vs compute-bound explanation; current product-page cluster model sizes
Version 13 · Oct 31, 2026, 11:53 PM
- Update: added a section on NVIDIA's October 2026 MoE inference case study (Humming kernels, online FP8 LM head, MTP speculative decoding) with results, explained the terms, and refreshed DGX OS details from the October 2026 release notes
Version 12 · Oct 29, 2026, 09:33 PM
- Update: note the Atlas inference engine, built for DGX Spark (GB10) and AMD Strix Halo, in the CUDA section.
Version 11 · Oct 25, 2026, 05:44 AM
- Correction: fixed GB10 memory speed, two-unit model size, Mac and AMD comparison specs, benchmark figures and misattributed quotes; added 2026 prices, 64 GB model, clustering
Version 10 · Oct 15, 2026, 10:11 PM