overview

Overview

⚠️ Sample. An AI-generated demo of LLM Wiki Newsroom — the "open source AI" topic is just the example corpus.

This wiki is a knowledge base comprising 4 source documents (2024~2025), 5 entities, 6 concepts, 3 field overviews, 1 analysis reports, 0 associative trails, and 0 timelines.

Sources are automatically classified into 3 topic clusters via Leiden topology clustering: Open-Source AI Definition(4), Open Weights(1), Licensing & Open-Washing(3). A single source may span multiple clusters (listed in every catalog where its weight is ≥0.3); for the full cluster list and members, see index or graph/_clusters.json.

This wiki maps the debate over what "open source" should mean for AI systems. The corpus gathers the OpenSourceInitiative's 2024 attempt to fix a definition, the endorsement and criticism it drew, and the looser "open weights" releases that dominate the market in practice. Three groupings organize the field: the formal definition and its data dispute, the weights-only middle ground, and the licensing terms and open-washing that surround both. The strongest evidence base sits with the definition grouping, where the OpenSourceInitiative is the recurring claimant across multiple sources.

Recent Changes

• 2025-05-19 — A legal primer frames the open-weights middle ground against the strict definition, using DeepSeek R1 as its leading example.
• 2025-01 — DeepSeek releases R1 with MIT-licensed weights but withheld training data, the defining open-weights case.
• 2024-12-05 — Criticism of the definition's omission of open training data is rounded up, led by the FreeSoftwareFoundation and OSI co-founder Bruce Perens.
• 2024-10-28 — The OpenSourceInitiative releases the Open Source AI Definition 1.0; Mozilla endorses it the same day, and Meta's Llama is judged non-compliant.

1. Open-Source AI Definition

On 2024-10-28 the OpenSourceInitiative released the Open Source AI Definition 1.0, the first binary standard for open AI: a system either grants the four freedoms — use, study, modify, share — or it does not, and qualifying requires data information, source code, and parameters. OSI validated Pythia, OLMo, Amber, CrystalCoder, and T5, and judged Llama non-compliant.

The grouping's actors split three ways. The OpenSourceInitiative is the standard-setter; Mozilla is the prominent endorser, framing openness as an AI-safety precondition; the FreeSoftwareFoundation is the dissenter building a stricter alternative. The anchor concept is OpenSourceAI, and the contested component is TrainingData.

The central collision is whether a workable binary standard or a maximalist open-data requirement should define "open." OSI and Mozilla argue an imperfect-but-clear definition serves developers and regulators now; the FreeSoftwareFoundation holds that without raw training data the label is hollow. The procedural fact that a 10-person board, not the full membership, approved the definition feeds the dispute.

Details: the Open-Source AI Definition field.

2. Open Weights

DeepSeek's R1, released 2025-01 under the MIT license with public weights but a withheld training corpus, is the defining case of the weights-only middle ground. It drew attention partly on a roughly $6 million training-cost claim that made a cheap-to-adapt release commercially compelling.

The grouping centers on DeepSeek as provider and on two concepts: OpenWeights, the posture of publishing parameters while keeping data secret, and FineTuning, the adaptation capability that makes such a release useful. It is defined by contrast with the fuller OpenSourceAI standard rather than by a dense actor roster.

The tension is practical adaptability against genuine transparency. Publishing weights lets users fine-tune without retraining, but the withheld data and algorithms mean the model cannot be fully audited, reproduced, or inspected for bias — which is precisely the partial release the strict definition was written to exclude.

Details: the Open Weights field.

3. Licensing & Open-Washing

This grouping covers the terms and rhetoric of "open" AI. Meta's Llama, released in early 2023 under a custom community license with usage restrictions, is the reference case the OpenSourceInitiative cites as non-compliant and as the canonical example of open-washing.

Licensing is where open-source AI diverges from software: permissive families (MIT, Apache) and copyleft families (GPL) were written for code, but a model adds data and weights that no software license covers. ModelLicensing and OpenWashing are the anchor concepts; Meta is the contested issuer, with OpenSourceInitiative judging and Mozilla advocating.

The collision is whether a familiar permissive license is sufficient signal of openness or whether component completeness is the real test. DeepSeek R1's MIT-licensed-but-data-less release shows the two are independent, and as "open source" enters regulation, a soft line lets restricted models claim benefits meant for open ones.

Details: the Licensing & Open-Washing field.

Cross-Domain Threads

One word, three standards. The whole corpus is a fight over a single phrase. The definition grouping wants "open source" to mean the four freedoms plus reproducible components; the open-weights grouping uses "open" to mean runnable-and-adaptable; the licensing grouping shows how the word gets attached to releases that satisfy neither. The OpenSourceInitiative's binary definition is the attempt to collapse these three readings into one, and the resistance to it — from the FreeSoftwareFoundation on one flank and from market practice on the other — is what keeps the field unsettled.

Training data as the fault line. TrainingData is the component that recurs across all three groupings. The definition grouping fights over whether it must be released, the open-weights grouping is defined by withholding it, and the licensing grouping shows that a permissive license says nothing about it. Whether "open" requires open data is therefore the single question that, once answered, would resolve most of the surrounding disputes at once.