Publication status¶

Available, Preview or RDI. This says whether your hardware and software can be bought today, and is independent of which division you submitted in.

The policy repository wins

This page summarises policy in submitter language. It does not restate it. Where this page and mlcommons/endpoints_policies disagree, the policy repository is correct and this page is a bug. Please report it.

Category	Hardware	Software
Available	All performance-determining components pass the four-point test at submission time	Available software stack
Preview	Not yet Available; you commit to Available within 180 days	Available, except software supporting substantially new hardware
RDI	Meets neither	Meets neither

RDI has a cooling-off period

An RDI component may not be submitted as Available or Preview until the later of two cohorts after the RDI submission, or 221 days after first publication as RDI. This prevents RDI being used to pre-publish on unavailable hardware and immediately reclassify.

Available: the four-point test¶

Every component that substantially determines ML performance must meet all four at submission time:

#	Criterion	Requirement
1	Pricing	Publicly advertised or available on request
2	Shipment	Shipped to or rented by at least one third party
3	Public evidence	Externally verifiable evidence it is shipping or purchasable today
4	Reasonably available	Purchasable or rentable by additional third parties by the submission date

Availability is judged on the balance of evidence. No single data point settles it. Uneven availability across geographies or channels is not penalised, provided availability exists in at least one market.

Proof of shipment, not announcement¶

These are not sufficient on their own

Announcements at a conference, trade show or keynote
Product listings on a marketing or "coming soon" page
Soft-launch press releases describing a future availability date
A "Preview" or "coming soon" entry in a cloud provider's catalogue
A stated or anticipated launch date without proof of current availability

Sufficient evidence:

A vendor product or pricing page with a live, functional call to action for qualifying customers — "Buy now", "Order", "Get started" or equivalent
Public vendor claims that the product is "now shipping", "generally available" or "in production", in a press release, earnings call or official blog post
Publicly verifiable fulfilment data — a cloud instance type in live pricing and availability APIs, orderable by any qualifying customer
Third-party review units or press loaners confirmed by the recipient
Public purchase orders or shipment confirmations in securities filings

For component vendors (accelerators, ASICs, memory, networking) the test is applied at the component's position in the supply chain: available to ship to system integrators, OEMs and ODMs under standard commercial terms.

Your own GA definition also binds you

If your organisation has a formal internal process defining general availability ("open shipping status", "phase 4 exit", "FCS") the submitted system must satisfy that internal definition in addition to every criterion above.

Reasonably available¶

Non-discrimination. Competitors must not be blocked. All software, drivers, firmware and support made available to any customer must be available to any other qualifying customer, including direct competitors, on equivalent terms.
Access conditions. Conditions common to generally available products are fine: financial qualification, customer size, support burden, export restrictions. "Early access" approval requirements are not.
Supply and lead times. Subject to market conditions, and not governed by these rules. Reproducing an MLPerf result gives you no priority slot or allocation right. A reviewer or auditor faces the same supply constraints as any other customer.

Qualifying pre-submission rentals or purchases are allowed to have been made under "early access" restrictions.

Available software stack¶

Category	Rule
ML frameworks (PyTorch, JAX)	Any commit in an official public repository. Open PRs adding architecture support are fine if publicly accessible
Inference servers (TensorRT-LLM, vLLM, SGLang, Triton)	Open-source: any public commit plus accessible open PRs. Commercial: official release or labelled beta in the formal release sequence
Accelerator compute libraries (cuDNN, NCCL, ROCm HIP)	Official release or publicly available labelled beta. One-off private binary drops do not qualify
Hardware drivers	Official release or public beta, downloadable through a standard vendor channel at submission time
Quantization tools	As frameworks if open-source; as compute libraries if closed-source binary
Custom OS / kernel patches	Exempt if unmodified commodity software. Patches that substantially affect ML performance must be upstream-merged or in a publicly accessible PR. Private patches are not permitted
Your own submission code and custom kernels	Not part of the "stack" — must be in your submitted source package

Labelled beta qualifies only if: the vendor has publicly committed the optimisations will ship in a future official release; the beta is a standard step in the release process, not a one-off private engagement; and it is publicly downloadable at submission time. A release candidate qualifies; an NDA engineering sample does not.

Software must be available at the time of submission. If a component becomes unavailable between submission and publication, tell the review committee. They may let it proceed if an equivalent public release exists, or require reclassification.

Available models¶

Weights and tokenizer must be available through commercially accessible channels. Weights under a licence prohibiting commercial deployment don't qualify, and must be submitted as Preview or RDI.

Division-specific additions¶

Standardized CoN — the endpoint URL must stay accessible for at least 90 days after publication, with a point of contact for access requests.

Serviced — the endpoint must be at full GA tier. Not "Public Preview", not "Open Beta", not "Generally Available in Preview", even if publicly accessible. Where your own GA definition distinguishes preview-accessible from production-GA, production-GA is required. The terms of service must permit third-party benchmarking; any MLCommons member in good standing must be able to access the endpoint under standard terms and attempt reproduction.

RDI division — always carries RDI status and cannot qualify as Available or Preview. If the system later becomes commercially available, make a new Standardized or Serviced submission.

Preview, and its clock¶

A Preview system is not Available at submission, but you commit to making it Available within 180 days of first publication and to re-submitting as Available at that point.


Clock starts	The date of the cohort the result first appears in — not the submission date
Duration	180 calendar days
Anchor	First publication. Corrections and amendments do not reset it

Results publish tagged "Preview — Available by [date]".

Declaring Preview¶

Set "availability_status": "preview" in the system description
State a "target_availability_date" within the window, as an ISO 8601 date
Identify with specificity which components are not yet available — "The X100 GPU is expected to begin customer shipments in Q3 2026". Vague statements like "targeting H2 2026" are not sufficient

Software waiver¶

The Available software stack requirement is waived for software supporting newly developed hardware that substantially determines ML performance. "Newly developed" means the hardware was not Available as of the previous cohort and was not submitted as Preview in that cohort. All other components must still meet the Available requirements.

Performance continuity¶

When Preview transitions to Available, the re-submitted result must be equal or better.

Throughput — system_tps and tps_per_user may each degrade by up to 5%, applied independently, to account for endpoints-workload variance.
Latency — the 5% tolerance does not apply to ttft_p90_ms or any latency percentile. Until a comparison method is ratified, a transition is not blocked on latency alone; a material unexplained latency regression is raised as an objection during the re-submission's own review.

Expiry and extensions¶

At the end of the 180 days:

Situation	Outcome
Available re-submission published	Preview result is superseded
No re-submission, no extension	Preview result is invalidated and removed at the next cohort — not archived
Approved extension in force	Remains Preview for the extended period

One extension only

A single extension of up to 60 calendar days may be granted by the review chair. It must be requested at least 30 days before expiry, with a written explanation and a revised, specific date. A second extension is denied and the result is invalidated. You can re-submit under RDI.

MLCommons maintains a public Preview Availability Tracker listing active Preview results, their first publication dates, target dates and days remaining, updated each cohort.

Transitioning to Available¶

Notify the review committee via a GitHub issue on the submission thread
Make a new Available submission with "availability_status": "available" and an "availability_url" pointing at a public product or ordering page
It goes through standard compliance and peer review
If the hardware or software configuration changed materially, you must re-run. Relabelling an existing Preview result as Available without re-running is not permitted where the configuration changed
On success the Available result publishes and the Preview result is retired

Referencing Preview results¶

All external references must use the qualified designation "MLPerf Endpoints Preview." Omitting the qualifier when referencing an unfinalized result violates MLCommons usage guidelines. The standard footnote applies.

Custom SKUs¶

Vendors, OEMs and ODMs may qualify custom SKUs as Available, including silicon variants produced exclusively at scale for one large-volume customer, if all three hold:

The SKU has shipped to at least one customer
It is available to purchase by similar customers at a volume the vendor determines, with customizations available on request where warranted
The differences from any standard SKU are clearly and publicly documented by the vendor

Hardware no comparable customer can order

Silicon in production at a hyperscaler but not orderable by any comparable customer fails criterion 2, and may have no GA commitment so can't be Preview, yet it isn't a prototype either. The working group is weighing a new "Production" tier against the current default. Until decided, it defaults to RDI, with the 221-day cooling-off attached. Tracked upstream as [CUSTOM-SKU].

These rules are a working draft

MLPerf Endpoints v1.0 rules are on an active development branch. Sections marked [TENTATIVE], [WIP], [WG Open Item] or [WG Decision Required] upstream can still change. Before committing hardware time to a submission, read Open questions and WIP rules.

Last verified against: mlcommons/endpoints_policies@v1.0_rules_dev (a7ec3cc), 2026-09-19.