CoreWeave Publishes MLPerf Results, Claims Cloud-Provider Leads

CoreWeave has published MLPerf Inference v6.1 results that it says put its NVIDIA GB300 NVL72 systems ahead of other cloud-provider submissions on selected multimodal and language-model workloads. The company is also arguing that those results came from production clusters, not a system tuned only for a benchmark.

CoreWeave entered four NVIDIA platforms and four model families in MLPerf’s Datacenter Closed, Available category. Its strongest provider-specific claims concern the GB300 NVL72, a 72-GPU rack system: the company reported 1,196 queries a second on Qwen3-VL-235B-A22B’s server test, which it called the highest among cloud providers.

For Llama 2 70B, CoreWeave reported 944,902 tokens a second in the server scenario and 1,136,100 in the offline scenario. It described both as the highest throughput among cloud-provider submissions using GB300 NVL72.

The GPT-OSS-120B claim is broader than the provider comparisons. CoreWeave reported 16,635 tokens a second per GPU offline and 16,118 in the server scenario from one GB300 NVL72 rack, calling both the highest per-GPU throughput of any v6.1 Datacenter Closed submission.

That ranking carries a material qualification. CoreWeave calculates per-GPU throughput by dividing a system’s total throughput by its number of accelerators. It says MLCommons did not verify that derived metric, so the per-GPU ranking has a different verification status from the underlying benchmark submission.

CoreWeave also reported a 19.8% increase in derived per-GPU server throughput for DeepSeek-R1-671B against its earlier MLPerf v6.0 GB200 NVL72 result. The submissions used different GPU counts—72 in v6.1 and 64 in v6.0—so CoreWeave normalized the comparison per GPU rather than presenting it as a total-system score.

CoreWeave says the submissions used production clusters, images and operational software available to customers, rather than a benchmark-specific tuned cluster. That makes the company’s case about more than a leaderboard position: it is presenting the measurements as evidence that its operating stack can deliver those results outside a dedicated test environment.