📨 AI for Social Impact Deep Dive: Open Weight Models

✍🏼 A Note From the Editor

Lots of folks in the social sector are concerned about transparency when it comes to AI tools. This Deep Dive unpacks a topic that many social sector professionals probably don’t know much about: open weight models. We'll get into what these are, how it might impact your vendor choice, and what creating a custom AI model might involve.

🔓 APIs and Open Source — A Quick Primer

APIs: Most of the specialized AI tools you're using are likely built on top of one of the same three underlying engines: ChatGPT, Claude, or Gemini. AI tools access these models via API, which is a connection that lets one software product tap into another's capabilities. That means when you sign up for some vendors’ AI platforms, your data is often flowing through one of those companies' servers.

Open Source: In traditional software, "open source" means the underlying code is publicly available for anyone to view, use, modify, and distribute. The opposite is proprietary software, meaning that its underlying code cannot be viewed or modified. Linux, WordPress, and Firefox are classic examples of open source software. The philosophy behind open source is that transparency and community collaboration produce better, more trustworthy software.

If we were to apply this concept to AI, it would mean that a true open source AI system would share not only the finished product, but also the training data, the code used to train it, and the full methodology behind its development. What we more commonly see are open weight AI models: systems where the trained model weights, the billions of numerical parameters that encode what the AI has "learned", are made publicly available, but the training data and process are not.

⚖️ Open-Weight Models: What They Are

The most prominent open-weight models you may have already heard about are Meta's Llama, Mistral (from a French AI company), and DeepSeek (from a Chinese AI company). These are large language models (LLMs), similar in capability to ChatGPT or Claude, but with one crucial difference: their weights (numerical parameters that encode what the AI has “learned”) are released publicly, so developers can download, run, and modify them. These models can be especially useful in highly regulated sectors that have strict requirements for in-house deployment and data sovereignty (think: healthcare, finance, etc.).

🦾 Open-Weight Models: Why Does This Matter?

The biggest implication is customization. With ChatGPT or Claude, you're working with whatever OpenAI or Anthropic has built. With an open-weight model, a technical team can train the model on your own data, so it speaks your organization's language, reflects your values, understands your context, and produces outputs tailored to your specific programs and communities.

So, how do you customize a model?

Fine-tuning is one option. When you fine-tune an open-weight model, you put an LLM through extra rounds of training, using data sets specific to a particular domain or organization, and embed this knowledge directly into the model.

Retrieval-Augmented Generation (RAG) supplements LLMs with current information, encoding organization-specific documents to augment training data. RAG is generally more resource-efficient and easier to update than fine-tuning, since it doesn't require retraining the model when your information changes.

✅ The Pros

Data Privacy and Sovereignty: Open-weight models running on your own infrastructure (or a trusted provider's) doesn't pass data through Big Tech's servers.

Values Alignment: Open-weight models can be fine-tuned to reflect your values. A model trained on advocacy communications will write differently than a general-purpose chatbot that's optimized to sound neutral on everything.

Cost Efficiency: Once a model is deployed on your infrastructure, per-query costs can be significantly lower than paying per API call to providers.

Reduced Vendor Lock-In: If your AI strategy is built entirely on one provider's API, you are at the mercy of their pricing and terms-of-service. Open weight models can give organizations more control over their AI stack.

⚠️ The Cons

Technical Know-How: Running and customizing these models requires technical capacity (read: machine learning expertise and ongoing maintenance). For the vast majority of social sector organizations, this is not a realistic in-house capability.

Opaque Training Data: As mentioned above, just because it’s open weight doesn’t mean it’s open source. Open weight models can still inherit biases from their training data and produce harmful outputs, requiring thoughtful testing before deployment.

Licensing Complexity: Not all open-weight models are freely usable for all purposes. Some models have commercial restrictions or usage thresholds that may apply to larger nonprofits or networks.

🗣️ Real Talk

Most organizations in the social sector aren't going to build a custom AI model anytime soon—and you don't need to. This deep dive is about helping you understand the broader landscape so you can ask better questions and evaluate vendors more confidently. Don't let the complexity stop you. Start with what you already have access to—Gemini in Google Workspace, Copilot in Microsoft, ChatGPT, Perplexity, or Claude—and keep learning.

👋🏼 About AI for Social Impact

I’m Joanna, and I’m on a mission to help folks in the social impact sector understand, experiment with, and responsibly adopt AI. We don’t have time to waste, but we also can’t get left behind.

Let’s move the sector forward together. 💫

♥️ Spread the Love

Spread the love and forward this newsletter to anyone who might benefit from a dose of AI inspo!