NVIDIA Unveils DGX Spark 64GB: A 1-PetaFLOP Grace Blackwell Desktop for Local AI Agents and Fine-Tuning
NVIDIA has announced a new 64GB configuration of its DGX Spark, a GB10-powered desktop AI system available from Acer, ASUS, Dell, Gigabyte, HP, and MSI. This configuration allows developers to start with a single system for local models and agents, then cluster two 64GB units to achieve 128GB of combined memory and additional compute as workloads grow.
The core message: run open models and always-on agents directly on your desk, then scale by clustering DGX Spark systems as needed—rather than relying on metered APIs.
The timing is strategic. Agents continuously consume tokens through tool calls, retries, long context windows, and multi-step planning. Token consumption has grown 14x since 2024, driven by the rise of autonomous AI agents and increasingly complex workflows. In 2026, with agentic AI becoming mainstream, local inference and fine-tuning on desktop hardware offer a cost-effective, private alternative to cloud-based APIs.
Key Features and Use Cases
- Local AI Agents: Run always-on agents without per-token costs or latency.
- Fine-Tuning: Train and adapt open models on your own data.
- Inference: Deploy models for real-time applications with up to 1 PetaFLOP of AI performance.
- Scalability: Cluster two units for 128GB memory and higher compute.
With the 64GB DGX Spark, NVIDIA targets developers and enterprises seeking to harness the power of Grace Blackwell architecture in a compact desktop form factor, ideal for the era of agentic AI.
via MarkTechPost