Don’t buy the DGX Spark: NVFP4 Still Missing After 6 Months

NVIDIA's DGX Spark, a $70,000 AI workstation designed for local AI development, is facing significant criticism from early adopters who report that a core promised feature remains undelivered more than six months after launch. The system was marketed with the key selling point of a complete, out-of-the-box software stack integrating Blackwell architecture with the NVFP4 (Neural Video Frame Processor 4), positioning it as a premium, finished product. However, users like the author, who owns two units, state that achieving functional NVFP4 support currently requires extensive manual configuration, backend switching, testing of community builds, and setting experimental flags—a far cry from the polished experience advertised.

This gap between marketing and reality has created a major pain point for professionals who invested in the DGX Spark for its promised capabilities. The hardware itself shows potential, but the overall user experience is described as immature and unstable, more akin to an experimental developer kit than a supported enterprise system. The author argues that NVIDIA appears to have pushed the product story before the underlying software was actually ready, leaving customers with a compromised system that's difficult to justify given its inherent bandwidth limitations. The core takeaway for the AI community is clear: prospective buyers should not purchase the DGX Spark assuming NVFP4 is a delivered, mature feature.