As tech companies pour billions into artificial intelligence, a counterintuitive strategy is gaining traction: giving away the code. The move to open source AI models reflects growing pressure to rein in development costs while accelerating breakthroughs in a field where talent and computing power come with a hefty price tag.
The AI landscape is shifting as companies like IBM, Meta and Mistral AI embrace open-source development, making sophisticated AI technology freely available to developers and researchers. While this approach could democratize access and accelerate breakthroughs through collaboration, critics warn that it could allow bad actors to misuse or modify the technology for harmful purposes.
Proponents say that open-source AI development makes it possible for researchers and developers around the globe to inspect, modify and improve the technology. This collaborative approach has already yielded significant advances in model efficiency. For instance, IBM's new Granite 3.0 AI models achieve performance comparable to larger systems while requiring just a fraction of the computing resources.
IBM is releasing its Granite 3.0 models as open-source software under the Apache 2.0 license, taking a markedly different approach from developers that keep their AI systems private. The company built its models using public datasets like GitHub Code Clean and StarCoder, allowing them to steer clear of the copyright issues that have led to lawsuits against AI companies that train their models on protected content, such as News Corp's current case against Perplexity.
The release includes 8B and 2B language models focused on enterprise tasks like retrieval augmented generation and classification, along with specialized variants for instruction and security monitoring. Supporting 116 programming languages and trained on 3-4 terabytes of tokens, the models are available through multiple platforms, including Hugging Face, GitHub and IBM watsonx.ai. The models range from 3–34 billion parameters and can be used for research and commercial applications without restrictions in nominally "open-source" releases.
"Having that diversity of thought and contributing as part of this open ecosystem is just a much more exciting proposition than keeping our models closed in a box," says Kate Soule, Program Director of Data and Model Factory at IBM Research. “We want the community to use it.”
The open-source movement has also gained momentum in Europe, where Mistral AI has emerged as an industry leader. The Paris-based startup has released increasingly capable models that developers can freely download and modify.
The open-source strategy creates a two-way exchange: Companies share AI models that only deep pockets could build while gaining insights from thousands of developers as they find novel uses for the technology. Restricting access, many now argue, means missing out on this collective innovation.
“Imagine there exists some small, novel tweak to your model architecture that, even leaving everything else unchanged, would significantly improve overall performance,” says Dave Bergmann, a Senior Writer at IBM Think. “If you release your model as only open weight and decline to disclose information and code for its architecture, you might never realize the opportunity there. But if 20,000 people mess around with your model code, someone will spot it.”
Meta's release of its open-source Llama models was a significant turning point. It meant that one of tech's biggest companies gave researchers and smaller organizations access to advanced AI technology they could never afford to build independently.
“We fundamentally believe that open-source AI is good for everyone, starting with developers who can take these models and fine-tune it, train it, distill it and use it wherever they want,” says Amit Sangani, Senior Director of AI Partner Engineering at Meta, during a recent AI Alliance meetup. “The more open-source AI is, the more transparent and safer it becomes. It gets widely scrutinized, and that’s a good thing—because when issues are found, people, including Meta, will fix them.”
Despite the advantages, open-sourcing AI technology could come with risks. Some observers say that open-source AI models can be more easily modified for harmful purposes, such as generating disinformation or creating malware. The lack of centralized control means that even if the original developers adhere to high ethical standards, modified versions of their models might not.
Nonetheless, the efficiency gains demonstrated by models like Granite 3.0 point to a potential future where AI capabilities become more accessible to organizations with limited resources. Small businesses that can’t afford to run massive language models could deploy smaller, more efficient, open-source alternatives tailored to their specific needs.
“Open source is going to enable more innovation in more places,” says Dave Nielsen, AI Alliance Community Director. “And that's just going to generate more excitement for more entrepreneurs and more customers, and it will create more revenue.”