Data API

Read-only JSON and CSV. No authentication required.

Base URL: https://kiankyars.github.io/model-tracker/api/v1/

Endpoints

• index.json: Endpoint discovery and available partitions.
• models.json: Complete catalog, metadata, and coverage information. Records are in the models array.
• stats.json: Catalog counts and recorded-release aggregates.
• candidates.json: Discovery candidates and review decisions; unverified entries are not catalog models.
• coverage.json: Historical source review by month, outstanding candidates, and last review dates. Reviewed months are not claims of exhaustive coverage.
• months/YYYY-MM.json: Records released in one calendar month.
• developers/DEVELOPER_ID.json: Records from one model developer.
• models/SLUG.json: One record. Read slug from the catalog; do not construct it from a display name.
• models.csv: Tabular download of the catalog.
• JSON Schema: Model record schema.
• OpenAPI: Machine-readable HTTP API description.

Only partitions listed in the index exist. An unpublished month, developer, or model path returns HTTP 404. Follow the index instead of guessing identifiers.

Querying

GitHub Pages serves static resources. Query strings do not filter API responses. Use a month or developer partition to reduce transfer, or fetch the full catalog and filter locally. The website's URL parameters are browser filters, not API parameters.

const base = 'https://kiankyars.github.io/model-tracker/api/v1/';
const response = await fetch(`${base}models.json`);
if (!response.ok) throw new Error(`HTTP ${response.status}`);
const { models } = await response.json();
const results = models.filter(model =>
  model.open_weights === true &&
  model.parameters_b !== null &&
  model.parameters_b <= 8 &&
  model.release_date?.startsWith('2025-')
);
console.table(results.map(({ name, release_date, parameters_b }) =>
  ({ name, release_date, parameters_b })
));

curl -fsSL https://kiankyars.github.io/model-tracker/api/v1/models.json \
  | jq '.models[] | select(.developer_id == "anthropic")'

Interpreting fields

• id: Stable canonical model identifier, including developer namespace.
• slug: Filename-safe identifier for the individual model endpoint.
• developer, developer_id: Model creator, independent of hosting provider.
• announcement_date: Official announcement day or month, when verified.
• availability: Verified release stage reached: announced, limited, or public. This does not assert that a historical service still runs. Missing or null means not recorded. Public release does not imply open weights.
• release_date: First verified external availability, including limited access. ISO day or month, or null; respect date_precision. Announcements have no release date until external access is verified. A released model can still have an unresolved date.
• open_weights: true when downloadable weights are publicly available for the model; false when no public weight release is available at the latest review. This is not an open-source license classification. A hosted variant can be closed even when its base model has open weights. Clients should tolerate null for records awaiting review.
• parameters_b: Reported model parameters in billions, or null. Normally total model parameters; parameters_scope identifies a reported subset, such as backbone or language_decoder.
• parameters_lower_bound_b, parameters_lower_bound_exclusive: A published lower bound when the point count is null; true means strictly greater, false means at least. Bounds display in the table and qualify for the 100B+ filter when at least 100B, but are excluded from point charts and known_parameters statistics. Size sorting uses the published bound.
• active_parameters_b: Active parameters in billions when known. active_parameters_scope identifies the measured component or inference phase, such as language_decoder or decode.
• parameter_components_b, active_parameters_by_phase_b: Optional named component and phase counts, in billions. Do not add rounded components to infer a complete checkpoint count.
• field_status: Optional explanations for unresolved fields: undisclosed, not_applicable, or unverified. audit_url links to the individual research report; audited_at dates that review.
• parameters_source: Where the size came from. Name-derived sizes are not independently measured.
• context_window: Maximum context tokens in the model catalog, or null.
• modalities: Input and output modality arrays.
• reasoning, tool_call: Source-reported capability flags; unknown values remain null.
• source_url, sources: Provenance for the record and source-backed corrections.
• upstream_present: Whether a record is in the current upstream snapshot. Removed models are retained in the historical catalog.

A model keeps the same canonical id as it moves from announcement to availability. Missing dates or sizes do not exclude a verified model from the catalog.

total_models and developer count include all catalog records. recorded_releases, by_month, and developer recorded_releases require a release date and exclude announcements. unknown_release_dates excludes models still marked announced. The availability counts report explicit reviewed states and not_recorded separately. Other top-level counts describe the complete catalog. Months with no recorded releases have no API partition; the UI fills empty months with zero. Neither zero nor a high recent count measures completeness.

The website date filter and date sort use the release date, falling back to the announcement date. All charts and the summary beneath them count only recorded releases.

Models can have overlapping types: a model may be both multimodal and reasoning-capable. Unknown numerical values are not zero and should be excluded from numerical averages. Use the catalog's coverage counts to disclose missingness in your own analysis.

Versioning and refreshes

The /v1/ contract supports additive fields. Clients should ignore unfamiliar fields and tolerate nulls. Data corrections can change historical aggregates; pin a repository commit when you need a reproducible dataset.

Releases are researched daily. GitHub Pages publishes changes after validation. updated_at is the later of source retrieval and completed research. source.retrieved_at dates the Models.dev snapshot; optional research.reviewed_at dates the latest research review.

Cache downloads and use conditional requests with ETag or Last-Modified when provided. GitHub Pages caching and hosting limits apply.

For agents

llms.txt links to data and documentation. Automated reading and reuse are permitted under the MIT license. Model weights and third-party publications retain their own licenses. See coverage.