'Open-source AI' is the wrong term — where open-weight and open-source diverge
Contents
Llama and DeepSeek are often called "open-source AI," but strictly speaking they are "open-weight" — only the model weights can be downloaded. According to the Open Source Initiative (OSI), a model only qualifies as open source if the training data and code are also released, allowing third parties to reproduce it.
Background — why this topic matters now
Meta has consistently marketed its Llama series as "open source," but the OSI has repeated since Llama 2 launched in 2023 that "this license is not open source." Even after Llama 3.1 was released, the OSI reaffirmed that it "is not open source by any definition," and the Free Software Foundation (FSF) is reported to have reached the same conclusion.
This debate stems from confusion caused by borrowing a term — "open source" — that was already settled in the software industry and applying it directly to AI models. In software, "source" means source code, but what anyone can download from an AI model is the weights, not the training data or training code. To close this gap, the OSI released the "Open Source AI Definition (OSAID) 1.0" in 2024, and the topic has kept flaring up every few months since, including on Hacker News.
Key data & current state
- OSAID 1.0's three requirements: To be recognized as "open-source AI," a model must disclose all three of data information, code, and parameters under OSI-approved terms (OSI, 2024). Data information means explaining the source, collection, curation, labeling, and filtering of the entire training dataset in enough detail for a skilled developer to build a "substantially equivalent system" — it does not require releasing the dataset itself in full.
- Llama 3.x license verdict: The OSI found that the Llama 3.1 and 3.2 licenses fail to meet the open-source definition because they impose commercial-use restrictions (a separate agreement is required above 700 million monthly active users) and limit certain use cases (OSI, 2024).
- DeepSeek: The model weights and inference code are released under the MIT license, but the pretraining and instruction-tuning data and training code are not disclosed. The OSAI Index, which tracks open-source AI, classifies DeepSeek's models as "open-weight" for this reason (OSAI Index analysis).
- The exception, AllenAI OLMo 2: It releases not just the weights but also the training dataset (Dolma), training code, a reproducible training recipe, and intermediate training checkpoints, making it the case that comes closest to meeting the OSAID requirements (AllenAI).
Interpretation: Lining up these four cases shows that the label "open source" actually lumps together a wide spectrum, from releasing weights alone to releasing the entire pipeline. The more widely used a model is — like Llama or DeepSeek — the more likely its data and code are to stay closed.
In-depth analysis
1) What "genuinely open source" requires under OSAID
OSAID applies software's four freedoms — to use, study, modify, and share — directly to AI systems, and stipulates that exercising them requires access to the "preferred form for modification." With weights alone, you can run and fine-tune a model, but you cannot retrain it from scratch or verify bias or copyright issues in the training data. The OSI compares this gap to the difference between a compiled binary and source code in software.
2) Why neither Llama nor DeepSeek qualifies as "open source"
The two models fail for different reasons. Separate from not fully disclosing data and code, Llama's license itself includes a user-count-based commercial restriction and prohibitions on certain uses, so it falls short on the "free to use" requirement. DeepSeek, on the other hand, has a license (MIT) with almost no restrictions, but fails on the "study and reproduce" requirement because it never discloses its data and training code in the first place. Analysis: in other words, "is the license free" and "is it reproducible" are separate axes, and satisfying one doesn't prevent falling outside the open-source definition on the other.
3) Why open weights still matter
Releasing weights alone still carries real industrial value. For companies running a model on their own servers (on-premises), fine-tuning it on proprietary data, or building and distributing derivative models, weights alone are often enough. In debates on Hacker News and elsewhere, practitioners frequently take the pragmatic view that "whether I can run it on my own hardware and retain control matters more than the legal definition." Security researchers, however, counter that weight-only releases make it easier to strip safety guardrails, and that there's no way to recall the model once a problem surfaces downstream.
Implications in Korea
The term "open source" is used just as loosely in Korea. Naver announced in April 2025 that it released three lightweight HyperCLOVA X models (SEED, 3B/1.5B/0.5B) in a form available for commercial use. LG AI Research is also reported to have uploaded K-EXAONE 2.0 (750B) to Hugging Face in July 2026, switching its license to Apache 2.0 so anyone can use it commercially. Both cases use the word "released," but the reported scope centers on weights and evaluation results — there is no confirmation that the training data or training code were also released. By OSAID's standard, these too are closer to the "open-weight" stage. For enterprises, this distinction matters in practice: checking license terms (scope of commercial use, redistribution limits) separately from whether the model is reproducible and verifiable helps avoid the kind of mistake where a team adopts a model assuming "it's open source, so mixing it in should be fine."
Outlook & variables to watch
- Positive factor: Fully open models that release data and code alongside weights, like the AllenAI OLMo series, keep emerging from academia and nonprofits, so cases meeting the OSAID standard may gradually increase.
- Risk factor: As "open source" hardens into a marketing term, concerns about "open-source washing" — using the label regardless of actual disclosure scope — keep being raised.
- Checkpoint: When adopting or comparing models, it's safer to check "what the license terms restrict" and "whether the data and code were released" separately, rather than simply asking "is it open source?"
Conclusion
- "Open-source AI" and "open-weight" are distinct concepts with different disclosure scopes, and OSAID 1.0 requires disclosure of all three: data information, code, and parameters.
- Llama fails to meet the OSAID standard because of license restrictions, while DeepSeek fails because it withholds data and code — different reasons for the same shortfall.
- Weight-only releases still carry practical value, but in situations that require reproduction or verification, don't take the word "open source" at face value — check the actual license text and disclosure scope yourself.
References
Contents
Related posts
When Models Become Commodities, Where Does the Money Flow? — Moats in the AI Value Chain
Model prices keep falling while model companies' revenue surges, and 89% of firms use AI but only 37% feel a profit impact. This piece traces that gap across the AI value chain's layers.
Why does power consumption keep rising as efficiency improves — AI and the Jevons paradox
AI inference costs are falling by nearly half every quarter, yet data-center power-demand forecasts keep accelerating instead. Here's why the Jevons paradox, first observed in the coal era of 1865, is repeating in almost the same shape in the age of AI efficiency.
Predictive policing and social credit — democracies aren't so different either: the convergence of surveillance technology
The UK's Gangs Matrix and Oakland's predictive-policing algorithm differ from China's social credit system in justification and procedure, but at the level of technique — flagging risk groups from past data — they look alike. The line dividing surveillance technology is transparency and accountability, not political regime.
Why Can't AI Do Just This One Thing? — The Concept of the 'Jagged Frontier'
Why AI outperforms humans on some tasks but actually hurts performance on other, seemingly similar ones — explained through the Harvard-BCG field experiment and the theory of 'asymmetry of verification,' along with criteria for deciding how far to trust AI on a given task.