Cloud AI infrastructure provider CoreWeave has achieved a landmark in AI computing, training DeepSeek-V3 — one of the largest and most capable open-weight AI models ever created, with 671 billion parameters — in just 2.02 minutes. The record-breaking result was achieved using 8,192 NVIDIA GB300 NVL72 GPUs distributed across 2,048 nodes, representing the largest GB300 cluster ever submitted in an MLPerf benchmark.
The feat was announced as part of the MLPerf Training v6.0 results published by MLCommons, the AI benchmarking consortium. CoreWeave submitted three separate GB300 NVL72 configurations and delivered the fastest results across all closed and available-cloud submissions for what MLCommons designated the suite’s most demanding workload — a testament to the performance density and efficiency of the company’s purpose-built AI cloud infrastructure.
What makes this result particularly meaningful is that it was achieved on the same CoreWeave Cloud infrastructure that is commercially available to customers — not a specially configured one-off system. This means enterprises and AI researchers can access the same class of performance that enabled the world record, democratizing access to extreme-scale AI training capability.
MLPerf Training v6.0 introduced two new benchmarks based on Mixture-of-Experts (MoE) architectures — DeepSeek-V3 and GPT-OSS 20B — reflecting the industry’s shift toward sparse computation where models activate only a fraction of their total parameters for any given input. DeepSeek-V3 activates just 37 billion of its 671 billion parameters per token, achieving remarkable efficiency without sacrificing capability.
The round drew record participation with 24 submitting organizations and 95 unique systems utilizing 13 different hardware accelerators. Cloud submissions more than doubled compared to the previous benchmark round six months ago, reflecting the rapid expansion of AI cloud infrastructure globally and the growing number of organizations with the resources to compete at the frontier of AI training.
CoreWeave’s streak of MLPerf leadership underscores the competitive advantage of purpose-built AI cloud infrastructure over general-purpose cloud platforms. As AI training workloads become more demanding with each new generation of frontier models, the specialized clusters and networking architectures that CoreWeave has optimized are proving decisive in delivering the performance that AI developers need.
