Key Takeaways:
- Understanding the difference between open source and open weight AI.
- Examining DeepSeek vs. Meta Llama licensing terms.
- While open-source AI models provide greater accessibility, it remains crucial for contracts professionals to carefully examine the associated license agreement.
Open-source AI is having a moment. One of the latest names to dominate AI discussions is DeepSeek, a Chinese artificial intelligence company, whose open-source DeepSeek-R1 model has made waves in the tech industry. If you’re wondering what “open-source” means in the context of AI, this article is for you.
What Is DeepSeek?
DeepSeek is an AI model family designed for various natural language processing applications, from chatbots to content generation. What makes it particularly significant is its open-source nature, which allows researchers, developers, and businesses to experiment with, modify, and deploy the model without proprietary restrictions.
What Does “Open Source” Mean in AI?
When we talk about open source in AI, we’re referring to the idea that an AI model’s core components—such as its model architecture (how the AI is structured and trained), training code (the software that builds the AI), and model weights (the learned parameters that define the AI’s behavior)—are publicly available for anyone to use, modify, and redistribute. See DeepSeek’s Technical Paper.
How Open Source AI Differs from Proprietary AI
Unlike open-source models, proprietary AI models—such as OpenAI’s GPT-4 or Anthropic’s Claude—keep their architectures, code, and model weights private. These companies typically provide access to their models only through cloud-based APIs, ensuring they maintain control over usage, monetization, and improvements.
Because open-source AI models make their core components publicly available, users can access them through web-based platforms or download and run them locally on their own machines. Running models locally enhances data privacy, as no data is sent to external servers, and provides flexibility for model fine-tuning or customization. But this approach requires substantial computing power. Check with your engineering or IT teams to determine whether this is an option for your organization.
Open Source v. Open Weight
When evaluating AI models, it’s important to look beyond labels like “open” to understand what the licenses actually allow. While some models are fully open-source and grant broad usage rights, others provide access to certain components—most notably, model weights—while still imposing restrictions on modifications and commercial use. This distinction is key for legal and contracts professionals navigating AI contracts and compliance.
One example of this is the difference between DeepSeek’s R1 models and Meta’s Llama models. DeepSeek-R1 is released under the MIT License, granting users full freedom to use, modify, and redistribute the model. In contrast, Llama models are “open-weight,” meaning Meta allows access to model weights but restricts how they can be used, especially in commercial applications.
DeepSeek R-1 Uses MIT License
DeepSeek-R1 is released under the MIT License, one of the most permissive open-source licenses available. The MIT License allows:
- Free use: Anyone can use, copy, and modify the model for any purpose, including commercial applications.
- Redistribution: Users can share the model, whether modified or unchanged.
- Minimal restrictions: The only major requirement is including the original license and copyright notice when redistributing.
Meta’s Llama’s License
Unlike DeepSeek, Meta’s Llama models are not released under an open-source license. Instead, Llama models have their own custom licensing terms, which:
- Restrict commercial use: Businesses must request permission or obtain a license from Meta for certain use cases.
- Prohibit redistribution: Users cannot freely share or modify the model in the same way they can with MIT-licensed software.
- Include additional terms: The license imposes conditions that limit how the model can be deployed, especially for competitors or commercial AI services. (“You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Meta Llama 3 or derivative works thereof).”)
Significantly, users are permitted to use DeepSeek architecture, code, and weights as a launching point to build or improve an LLM, while users may not similarly do so with Llama.
Why Contracts Professionals Should Care
While open-source AI models provide greater accessibility, it remains crucial to carefully examine the associated license agreement. For instance, when using an open-source model through the web-based interface, the model provider may claim broad usage rights over your data inputs and outputs.
Learn More: For guidance on evaluating AI tools, refer to my Contract Nerds article on key considerations when onboarding an AI tool.
For legal and contracts professionals dealing with AI-related contracts, it’s important to recognize the distinctions between open-source models, open-weight models, and proprietary AI. This distinction matters because licensing affects how AI models can be incorporated into or used to develop new products and services.
Questions to consider include:
- What licensing terms apply to the AI model in use?
- Are there restrictions on commercial use?
- Are there attribution or notice requirements?
- What are the downstream obligations if modifications are made to the model? (For example, the open source General Public License requires any modifications/derivatives to also be subject to the GLP.)
DeepSeek’s rise highlights the increasing role of open-source AI in shaping the future of technology. While open-source AI offers transparency and flexibility, it also introduces legal considerations that cannot be overlooked. As AI adoption accelerates, understanding these nuances will be critical in drafting contracts and internal policies that balance innovation with compliance.
For more on AI and Contracts, check out my full column here.
