Skip to content

MLPerf Endpoints — Submitter Documentation

How to prepare, validate and submit an MLPerf® Endpoints benchmark result.

MLPerf Endpoints measures how well an inference endpoint serves generative AI models. Instead of one number, a result is a Pareto curve of throughput against per-user speed across concurrency levels. New to it? Read the four Understand pages, then follow the Submission workflow from Before you begin through step 9.

  • New here

    What the benchmark measures, how a submission moves through review, and which division you belong in.

    Start with the concepts

  • Ready to submit

    The nine-step path from an empty directory to a submission in the review queue.

    The submission workflow

  • Rules

    What you must meet, what gets you rejected, and which rules are still moving.

    Requirements

  • Reference

    Three CLIs, three disclosure files, one bundle layout, every compliance check.

    Look something up

Round status

Item Current value
Submission model Rolling — submit any day, no fixed deadline
Publication cohorts 1st and 3rd Wednesday monthly, 08:00 Pacific
Cohort cutoff Automated checks must pass ≥ 1 business day before the Wednesday
Rules version v1.0 (draft)
Milestone v1.0 submission round opens 2026-10-12, then rolling

There is no submission deadline. Submissions are timestamped on receipt and enter review immediately; results publish in the next cohort they clear (Submission Rules §4.1).

Coming from MLPerf Inference?

Six things work differently here:

MLPerf Inference MLPerf Endpoints
Schedule Fixed rounds, advance registration Rolling; register by holding a PRISM token
Result A score at one operating point A Pareto curve of 7–32 points, the Offline result included
Divisions Closed / Open Standardized / Serviced / RDI
Cross-query KV reuse Prohibited Permitted, with a per-query salt
TTFT percentile — P90 for v1.0 (v0.7 used P95)
Power Measured, optional Provisioned power, declared per system; required for Standardized

Full detail in What is MLPerf Endpoints? and Model equivalence.

Before you invest hardware time

The v1.0 rules are a working draft

MLPerf Endpoints v1.0 rules are on an active development branch, and several sections that determine how long your runs must be and what you must disclose are explicitly marked tentative or pending working-group ratification. Read Open questions and WIP rules before you commit accelerator time.

Authoritative sources

This site is a guided path over the sources below. It doesn't restate policy. Where this site and the policy repository disagree, the policy repository is correct.