At GTC 2026 Berlin, announced that DGX Spark platforms now support 8-way clusters, doubling the previous 4-way limit. This upgrade enables 128 GB systems to reach 1 TB of memory, while 64 GB configurations provide 512 GB for intensive AI workloads.
The company also launched the 74B 24GB model and Recursive Flint, a 40B MoE coding model optimized for RTX Spark. New RAM-efficient checkpoints were released for models like Qwen3.8-Flash-Next and GLM-5.3 to improve local hardware utility.
AI inference performance has increased by 40% on DGX and RTX systems. With 1 TB of memory, DGX Spark can now run models like GLM 5.3, while the delivers a 40% uplift in Qwen 3.8 Flash-Next, significantly boosting local AI capabilities.