Redistributed

Meta's Llama is often called "open source." It isn't — not in the traditional software sense (Widder et al., 2023). Understanding the distinction matters for who actually controls AI infrastructure.

What "Open Source" Traditionally Means

In software, open source has a specific definition per the Open Source Initiative (OSI, 1999):

Source code is available — you can read how it works
Free redistribution — you can share it without paying
Derived works allowed — you can modify it
No discrimination — anyone can use it for any purpose

Open source software gave us Linux, Firefox, and most of the internet's infrastructure. It worked because software is reproducible — anyone with a computer can compile code.

What "Open Weights" Actually Means

Models like Llama release:

Model weights — the numerical parameters that define the model
Inference code — software to run the model
Sometimes training code — software used to train (but not always)

They don't release:

Training data — what text the model learned from
Training infrastructure — the full recipe to reproduce the model
Data processing pipelines — how raw data became training data
Human feedback data — the ratings that shaped behavior via RLHF
Internal evaluations — how the company assessed the model

This is better than keeping everything proprietary. But it's not open source in the traditional sense.

Why the Distinction Matters

Reproducibility. With traditional open source, you can rebuild the software from scratch. With open weights, you can run the model but not retrain it. You're dependent on whatever decisions the original trainers made.

Verification. With source code, you can audit what the software does. With model weights, you can probe behavior empirically but not fully understand why. The "logic" is distributed across billions of parameters without interpretable structure.

Training data concerns. Copyrighted material, personal data, and biased content may be in the training set. Without knowing what's there, downstream users can't assess legal or ethical risk.

Modification limits. You can fine-tune an open-weights model, but you can't fundamentally change what it learned in pre-training. If the base model has problematic knowledge or tendencies, you're working around them, not removing them.

The Business Model

Why do companies release open weights when training cost tens of millions?

Ecosystem lock-in. If developers build on Llama, they're part of Meta's ecosystem. When Meta releases the next version, switching costs are real.

Compute arbitrage. Open weights means everyone else bears the inference costs. Meta spent once on training; the community spends continuously on running.

Standards setting. If Llama becomes the default base model, Meta shapes the infrastructure without paying the full cost of deployment.

Regulatory positioning. "We're open" is a defense against regulation. It's also a way to make regulation harder — if the model is everywhere, controlling it is difficult.

What True Open Source AI Would Require

For AI models to be open source in the traditional sense, releases would need:

Training data or at least detailed provenance documentation
Complete training code and infrastructure specifications
Data processing pipelines
Human feedback data
Evaluation benchmarks and results
Modification rights without use restrictions

Some projects are attempting this (BLOOM, OpenLLaMa, RedPajama, and AI2's OLMo). They're smaller than frontier models, because truly open development can't match the concentrated resources of big labs. But the gap is narrowing: AI2's OLMo 3, released in 2025, is the first fully open 32B reasoning model — weights, training code, data, and evaluation suite all under Apache 2.0. EleutherAI's work on open datasets and training methodology has moved the entire field toward transparency. These projects demonstrate that genuine openness is viable, even if it cannot yet match frontier scale.

The Infrastructure Implication

The open-weights wave is better than pure proprietary control — substantially better. It has enabled a thriving ecosystem of small, efficient models, local inference tools like Ollama, and fine-tuned applications that would be impossible under a fully proprietary regime. Credit is due to Meta, Mistral, and others for releasing weights that genuinely expanded access. But it's not the full democratization some claim.

The companies that can train frontier models still control the frontier. Open weights let others build on top but not catch up. The underlying infrastructure — compute, data, human feedback — remains concentrated.

Understanding this distinction matters for policy. "Open source" AI shouldn't get the regulatory carve-outs that open source software gets unless it meets the same standards of transparency and reproducibility. Right now, it doesn't.

Sources

Open Source Initiative. "The Open Source Definition." 1999. OSI
Widder, David Gray et al. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI." SSRN, 2023. SSRN
Kapoor, Sayash et al. "On the Societal Impact of Open Foundation Models." 2024. arXiv
Liesenfeld, Andreas et al. "Opening Up ChatGPT: Tracking Openness, Transparency, and Accountability in Instruction-Tuned Text Generators." CUI, 2023. ACM

General

Open Source vs. Open Weights