AI 基礎設施
MLPerf Endpoints 0.7 Moves Cloud Inference to Continuous Submissions, but Initial Release Still Lacks Cost Normalization
MLCommons has launched Endpoints 0.7 for managed inference services, replacing twice-yearly static comparisons with automated submissions, continuous review, and interactive charts. The initial release covers five providers and three benchmarks, but cost normalization, agentic workloads, and broader participation will not arrive until version 1.0.
MLCommons released [MLPerf Endpoints 0.7](https://mlcommons.org/2026/07/mlperf-endpoints-v0-7-release/) on July 28, expanding the scope of evaluation from self-managed servers to the cloud and managed inference endpoints that enterprises actually purchase. Traditional MLPerf results are published in fixed rounds, making it difficult to keep pace with updates to models, accelerators, and service offerings. Endpoints therefore introduces an automated submission pipeline, continuous review tools, and a dynamically filterable results interface, with the goal of comparing the tradeoffs among latency, throughput, and service configurations across providers.
The initial 0.7 results come from CoreWeave, Google, Intel, KRAI, and NVIDIA and span three benchmarks. The underlying framework and rules are available in the [MLCommons Endpoints repository](https://github.com/mlcommons/endpoints). The significance of this design is not that it produces another “fastest GPU” leaderboard, but that it attempts to measure the complete endpoint: the same hardware paired with different batching strategies, inference engines, quantization methods, and service constraints can deliver dramatically different usable performance. If continuous submissions can maintain consistent review standards, the system should also be better suited to procurement comparisons among emerging cloud providers, hyperscalers, and specialized service providers.
However, version 0.7 is explicitly still a foundational release. MLCommons does not yet provide comprehensive cost or power normalization, and participants represent only a small share of the market. The current results therefore cannot directly answer operational questions such as tokens per dollar, stability under peak traffic, or cross-region latency. The organization plans to release version 1.0 later this year, adding agentic workloads and opening rolling submissions to more members. Engineering teams should next watch how agentic testing defines tool latency, success rates, and retry costs, as well as whether providers disclose enough information about model versions, parameters, and rate-limiting configurations to reproduce their results.