Capability
Open-Weight Share of Frontier Model Releases
Trailing 12-month share of Epoch AI frontier-model releases with downloadable weights under unrestricted, restricted-use, or non-commercial terms. It tracks access to frontier-scale model artifacts for research and deployment, without implying open source, broad model coverage, usage, or market share.
Open-Weight Share of Frontier Model Releases
Open-weight models accounted for 9.09% of Epoch frontier-model releases in the trailing 12 months ending 2025-07-31.
Methodology
The source is Epoch AI's downloadable frontier_ai_models.csv. Epoch defines a frontier model as a model that was in the top 10 by training compute as of its release date. Every one of the 137 rows in the downloaded file has Frontier model = True. This chart retains source rows with Publication date on or after 2023-01-01 and no later than the 2026-08-27 research cutoff; the newest retained source publication date is 2025-07-09.
Epoch provides two relevant fields in the frontier file: the categorical Model accessibility field and the derived Open model weights? field. A release is counted as open-weight when Model accessibility is Open weights (unrestricted), Open weights (restricted use), or Open weights (non-commercial), matching Epoch's published definition. As a data-quality cross-check, all 29 frontier rows published from 2023 onward have non-empty accessibility values, and this three-category classification agrees with Open model weights? = Yes for every row.
For month-end t, let S(t) be the first day of the month 11 months before t. The denominator D(t) is the number of retained Epoch frontier releases with S(t) ≤ Publication date ≤ t. The numerator O(t) is the number of those releases in one of the three open-weight accessibility categories. The plotted value is 100 × O(t) / D(t), rounded to six decimal places in the CSV. This is a 12-whole-calendar-month rolling share sampled at each month-end.
The first observation is December 2023, the first endpoint with a complete 12-calendar-month window beginning January 2023. The 12-month window is the conventional trailing-year view and is short enough to respond to changes in release policy without the unusual smoothing choice of an 18-month window. The 29 retained source releases include 1–4 releases in most months, so point-to-point movement remains sensitive to individual classifications.
The series stops at July 2025, the month containing the newest publication in Epoch's frontier download. It is not rolled mechanically through August 2026: doing so would age older releases out of the denominator while adding no newer source records, creating a misleading trend driven by the file's release-date cutoff. Epoch's downloads page marked the frontier file updated August 4, 2026, and the file was rechecked on August 27, 2026.
Open weights is not synonymous with open source. It means the trained parameters can be downloaded, potentially under use or commercial restrictions. This chart makes no claim that training code, inference code, training data, architecture details, or an Open Source Initiative-approved license are available. API access and hosted access without downloadable weights are counted in the denominator but not the numerator.
The 12-month window is the standard trailing-year convention and avoids presenting an unusual 18-month smoothing choice. Each denominator and numerator is recomputed from the dated frontier-release rows; no weights or unsupported aggregation are used.
Frequently asked questions
What does open weight mean in this chart?
It follows Epoch AI's accessibility taxonomy. A model counts when its weights are downloadable under an unrestricted, restricted-use, or non-commercial classification. API-only, hosted-only, and unreleased models do not count as open-weight.
Are open-weight models the same as open-source models?
No. Downloadable weights can come with significant use or commercial restrictions, and a release may omit training code, data, or other materials expected in open-source development. The chart deliberately says open weight rather than open source.
Why is the open-weight share relevant to the AI industry?
Downloadable weights can broaden who can inspect, adapt, self-host, and study frontier-scale models. The chart tracks the release mix of model artifacts, however, not downstream adoption, license freedom, deployment cost, safety, or the capability gap between open- and closed-weight systems.
Why use a 12-month rolling share?
Epoch's post-2023 frontier subset is small: individual quarters contain only 1–5 releases. A 12-month window retains the conventional trailing-year view while preserving monthly timing; the denominator remains small, so individual classifications can still move a point noticeably.
Why does the chart end in July 2025 when the source was checked in August 2026?
July 9, 2025 is the newest publication date in the frontier CSV downloaded and checked on August 27, 2026. Extending rolling windows past that month would create changes only because older releases left the window, not because the source added newer frontier releases.
Does this show the share of all AI model releases that have open weights?
No. The denominator is limited to Epoch models that were in the top 10 by training compute when released. It measures accessibility within that compute-frontier subset, not all model releases, usage, capability, or market share.