When a closed model returns a confident wrong answer, you have no path to a root cause: the training data is undisclosed and the weights inaccessible. That gap in visibility is the real dividing line between open and closed LLMs. Cost and capability vary across both categories; auditability does not.
What “open-source” formally requires
The Open Source Initiative published the Open Source AI Definition 1.0 (OSAID) in October 2024, the first consensus standard for what makes an AI system genuinely open. It grants four freedoms: to use, study, modify, and share the system without needing to ask permission. Exercising those freedoms in practice requires three components, each available under OSI-approved terms:
- Data information: enough detail about training data for a skilled person to build a substantially equivalent system
- Code: the complete training and inference source code
- Parameters: the model weights
A model that releases only weights, without training data documentation or source code, does not meet this standard. The OSI explicitly distinguishes “open weights” from “open-source AI,” noting that open weights enables deployment and fine-tuning but falls short of the freedom to study and reproduce (OSAID 1.0, Open Source Initiative, 2024).
What closed-source means in practice
Closed-source LLMs are accessible only through a provider API. The weights are proprietary and training details undisclosed; users submit prompts and receive outputs, and what happens between those two events is the provider’s prerogative. GPT-4 (OpenAI) and Claude (Anthropic) follow this model.
How they compare
| Axis | Open-source / open-weight | Closed-source |
|---|---|---|
| Weights access | Public (open-weight); public plus code and data (open-source) | None |
| Fine-tuning | On your own hardware, full control | Through provider API only, if offered |
| Data residency | Prompts stay in your own infrastructure | Processed on provider servers |
| Cost model | Compute costs you provision | Per-token API pricing |
| Licensing | Varies: MIT/Apache to custom commercial limits | Provider terms of service |
| Support | Community or commercial vendors | Provider SLA |
| Auditability | Weights inspectable; full reproducibility requires open-source | None |
The open-weight middle ground
Most widely used “open” models occupy a category the industry calls open-weight. Llama (Meta), Gemma (Google), Qwen (Alibaba), and Phi (Microsoft) release model parameters publicly. That is enough to run locally and fine-tune on custom data, with prompts staying in your own infrastructure. None of them release the full training data or code stack the OSAID requires.
A 2025 study examining more than 100 state-of-the-art models identified widespread “open-washing”: applying the open-source label while maintaining insufficient transparency, a practice the authors found undermines reproducibility and independent bias assessment (Sapkota et al., 2025, arXiv:2502.18505).
The OSAID lists fully compliant models: Pythia (EleutherAI), OLMo (AI2), and the LLM360 models Amber and CrystalCoder. These are the exception, not the norm.
What the distinction changes
For most deployments, the practical choice is between open-weight and closed-source. Open-weight covers the cases people most often cite: self-hosted data residency and fine-tuning on private data. For a broader look at the risks of tying an operation to a single closed provider, see what is AI vendor lock-in.
Closed-source models trade those controls for less operational overhead. A capable model is a few API calls away, with no infrastructure to provision or model serving to manage.
For guidance on choosing between them for a specific context, how to choose an AI model covers the decision variables in detail.
The non-obvious point: in common usage, “open-source LLM” almost always means open-weight. That gap between label and standard matters little for running a model. It matters significantly for reproducing one.
Sources: Open Source AI Definition 1.0, Open Source Initiative, October 2024; Sapkota et al., “Comprehensive Analysis of Transparency and Accessibility of ChatGPT, DeepSeek, and other SoTA Large Language Models”, arXiv:2502.18505, February 2025.
Questions, answered
Is Llama open-source?
Technically no, by the Open Source Initiative 2024 standard. Meta releases Llama weights under a custom license but not the full training data and code the OSAID requires. The accurate term is open-weight. You can run and fine-tune Llama, but you cannot fully reproduce it from training.
Can I fine-tune a closed-source model?
Some providers offer fine-tuning APIs. But the work happens inside the provider infrastructure, not on your own hardware, and the resulting model remains inaccessible to you as weights. The degree of control is narrower than with an open-weight model you run locally.
Is my data private with open-weight models?
Running an open-weight model on your own infrastructure means prompts and outputs never leave your environment. That is a genuine difference from a closed-source API, where the provider processes your data. How much it matters depends on what you are sending: general coding tasks differ from patient records or confidential contracts.
Which costs more?
Closed-source models charge per token via a provider API; cost scales with usage. Open-weight models require compute you provision yourself. At low to moderate volume, the API is usually cheaper. At high volume or strict latency requirements, self-hosted open-weight models can cost less per inference call.
Brief is a team of AI associates you direct in plain language. Opening to a small group at a time.
Request access