Which commercially usable models changed license evidence during the last 30 days?
OPEN-WEIGHT SEARCH + HISTORICAL EVIDENCE
The historical evidence graph for open-weight AI.
Search the current open-weight ecosystem, then inspect how verified facts change over time. Open Model Weights connects model records to primary sources, repository revisions, evidence snapshots and field-level diffs.
703model entities
710observed evidence snapshots
7retained field-change sets
2026-10-02/..temporal coverage
DIRECT ANSWER
What is the Open Model Weights Temporal Evidence Graph?
It is a time-aware, source-linked knowledge layer for open-weight AI. A normal model directory tells you the current value of a field. The Temporal Evidence Graph also records when Open Model Weights observed that value, which repository revision supported it, which primary sources were checked, and whether a later observation produced a material field change.
The model search remains the front door. The graph adds memory: current model discovery, historical evidence and machine-readable change intelligence become one system rather than separate pages.
SEARCH IS THE FRONT DOOR · HISTORY IS THE MEMORY
One product, four ways to interrogate it.
01020304
Search current models
Find field-verified open-weight models by developer, license, context, format, memory and other evidence-backed fields.
Search the registry →Inspect one model over time
Open the evidence ledger for a model to see observed snapshots, repository SHAs and retained diffs.
View an example history →See what changed today
Review daily verification results and source-derived additions, removals, releases and model changes.
Open today’s change ledger →Ask from an AI agent
Use versioned JSON and MCP so machines can retrieve current records, evidence history and temporal change data.
Open machine access →TEMPORAL DATA MODEL
Model → field → value → source → revision → observation → diff.
The graph distinguishes source facts, observed timestamps and derived values instead of flattening them into one confidence score.
01Observe
Check the source repository revision and primary evidence.
02Snapshot
Store normalized field evidence with repository SHA and evidence hash.
03Compare
Compare a new observation with the previous observed state.
04Retain
Keep material field changes instead of overwriting history.
05Connect
Link snapshots to developers, licenses, formats, runtimes and declared lineage.
WHAT IS TRACKED
Evidence that stays useful after the current page changes.
| Evidence group | Primary source | Temporal behavior |
|---|---|---|
| Weights | Repository API and exact artifact listing | Exact recognized files and formats can be compared between observed revisions. |
| License | Repository license files, model-card metadata and linked publisher terms | Changed terms or classifications create a new evidence state rather than silently replacing the old one. |
| Context & parameters | Structured config/API, then explicit source text | Observed specification changes remain attached to the revision where they were seen. |
| Formats & precision | Repository artifacts and dtype/config signals | Newly observed formats or precision artifacts can become dated change events. |
| Runtime signals | Publisher repository plus explicit primary runtime releases | Support signals remain evidence-scoped and do not become an independent runtime test. |
| Lineage | Declared base-model metadata | Parent relationships are retained only when the checked source declares them. |
| Training disclosure | Repository files and model-card disclosure | Disclosure state can change over time; undisclosed information is never backfilled by inference. |
| Source identity | Repository revision SHA and evidence hash | Each observation can be tied to the exact checked revision and normalized evidence projection. |
BUILT FOR AI AGENTS
Temporal questions become answerable.
Which models gained an officially observed GGUF or quantized artifact after a given date?
Which context-window or parameter claims changed between two observed repository revisions?
What did Open Model Weights actually observe for this model on a specific date, and which sources supported it?
IMPORTANT SEMANTICS
Historical evidence is not retroactive omniscience.
The graph only claims states that Open Model Weights actually observed. It does not reconstruct a publisher’s entire pre-observation history and it does not convert absent evidence into a negative claim.
Observed at
The time Open Model Weights captured the source-backed state, not automatically the publisher release date.
Repository SHA
The checked source revision when the repository exposes one.
Evidence hash
A stable hash of the normalized evidence projection used to detect material state changes.
Unknown stays unknown
Missing or undisclosed evidence remains explicit rather than being interpreted as false.
MACHINE-READABLE ACCESS
Use the same evidence layer from code and agents.
Temporal Evidence Graph JSONNodes, edges, observations and retained diffs →History indexSnapshot and diff counts for every model →Per-model ledgerFull observed evidence history for one model →Change feed JSONRolling human-meaningful registry events →OpenAPI 3.1Versioned JSON endpoint description →MCP for AI agentsSearch and retrieve evidence programmatically →
COMMON QUESTIONS
How the search engine and evidence graph fit together.
Is Open Model Weights still an open-weight model search engine?
Yes. Search and discovery remain a core product surface. The Temporal Evidence Graph adds historical provenance and change intelligence behind the searchable current records.
Does every daily check create a new snapshot?
No. Unchanged evidence does not need a duplicate historical state. A new snapshot is retained when the normalized evidence changes.
Can the graph prove what a publisher said before Open Model Weights started observing it?
No. History begins at the first observation retained by Open Model Weights. Earlier states are not invented.
Why is this useful for AI agents?
Agents need more than a current value: they benefit from source identity, temporal scope, explicit unknowns and machine-readable diffs when answering audit or change questions.