New research shows that open-weighted artificial intelligence (AI) models could save billions of dollars in costs, even though companies still generally prefer closed systems.

As AI systems expand outward from capital-intensive LLM systems controlled by a few companies, SLM systems are rapidly rising in popularity thanks to open-source systems, and their open weighting systems are quickly narrowing the performance gap with closed platforms.

In an interview with EE Times, Mozilla Foundation President Mark Surman said that the high costs required to train and run large AI models are driving the industry to seek more energy-efficient and low-cost innovation paths.

"The cost structure of open-source AI is very different from that of traditional open-source software because it involves computing resources, energy and infrastructure," Surman explained. "This year, the industry will focus more on small language models for specific application scenarios, while improving efficiency through distributed training and making use of idle computing resources."

Surman sees the Llama model released by Meta as a key turning point. He points out that before this, large language models seemed destined to be difficult to open source—almost all core technologies were firmly controlled by companies such as OpenAI and Anthropic.

"For most people, AI is chatbots or social media feeds that predict reading preferences. AI is everywhere, but its underlying architecture is like Lego bricks, made up of different components," Surman explained. "When users use chatbots or browse social media feeds, companies like OpenAI, Meta, or Alibaba have already assembled these bricks in their own way and packaged them into a 'box.' Open source means access to all the bricks."

Break the box

Surman points out that to determine whether AI is truly "open," one must examine each component of the system. He cites several models that have been made available to developers, including Gemma and GPT-OSS from major US tech companies, Qwen and Kimi from China, and a series of models from Mistral AI in France.

Distinguishing between open-weight models and fully open-source models is crucial. The pre-training process of open-weight models is a black box; developers can fine-tune and adapt the model, but cannot see its construction. Fully open-source models, on the other hand, allow developers to view the training data and checkpoints during pre-training. While fully open-source models exist, their performance still lags behind open-weight and closed-source models. Surman states that more investment is needed to improve the capabilities of fully open-source models.

[Editor's Note: Open weight models make the key parameters of an AI model, namely the weights, freely available to the public, allowing anyone to download, run, or fine-tune the model on their local device. However, unlike "fully open-source" models, open weight models typically do not disclose training data, training code, or complete architectural details.]

However, open-source AI is not limited to the models themselves. It also includes open-source datasets, open solutions for orchestrating computational infrastructure, and open architectures and coordination methods for building AI systems. "Technologies such as Transformer, large language models, reinforcement learning, and agent frameworks are already well-known. Open weighted models are gradually approaching the capabilities of proprietary models, sufficient to meet the needs of most users," Surman revealed.

At the same time, the academic community has begun to quantify the economic benefits of this transformation.

A "low estimate" of billions of dollars

In November 2025, Frank Nagle, a research scientist at MIT’s Digital Economy Project and chief scientist at the Linux Foundation, and Daniel Yue, an assistant professor in the field of information technology management at Georgia Institute of Technology’s Scherer School of Business, co-authored a paper that provided empirical support for Surman’s viewpoint.

They studied the LLM inference market—the stage where models are trained to generate outputs based on user instructions—which is currently the area with the highest concentration of daily AI spending. They found that, in terms of both price and performance, the utilization rate of open-weight models is far below their potential.

The study used data from OpenRouter, a platform that distributes API requests to dozens of inference service providers, covering 0.3% to 1.06% of the total market consumption. Researchers tracked daily token usage, price, and model supply from May to September 2025.

The research reveals a paradox in the AI economy: closed models from companies like OpenAI, Anthropic, and Google process 80% of the tokens and contribute over 95% of the revenue through proprietary APIs. In contrast, open-weighted models (whose weights are publicly available and hosted by any inference service provider) process 20% of the tokens but only generate about 4% of the revenue.

The price difference is even more pronounced. On average, the cost of an open-weight model is only 15.66% of that of a closed-weight model, meaning the latter costs about six times more. Surman explains that this difference stems from freedom: "Using open-source technology, you don't have to pay exorbitant fees, and you still enjoy flexibility and scalability."

The paper points out that open-source models are cheaper due to structural differences. Because any organization can host open-source models, numerous inference service providers compete to offer these services, driving prices close to marginal cost. In contrast, closed models are typically served by only one or two original development companies and their possible cloud partners, allowing them to maintain a higher premium.

Despite the significant price difference, the performance gap between the two is not substantial. Nagle and Yue's research infers that in mainstream benchmarks such as GPQA, MMLU Pro, LiveCodeBench, and LM Arena, open-weighted models achieve approximately 90% of the performance of closed-weighted models. Specifically, they score 89.6% on the GPQA graduate-level inference test. According to their research paper, the time required for open-source models to catch up with closed-weighted models is also decreasing: an average of 27 weeks in the first half of 2024, shortening to 17 weeks in the second half of 2024, and further shortening to 13 weeks in the first half of 2025.

Even after incorporating price and performance metrics into regression analysis, the actual usage rate of open-source models remains 63-88% lower than that of comparable closed models. The study also found that in several cases, closed models are not only more expensive but also perform worse than existing open-source alternatives, yet users continue to choose the former. The paper simulates scenarios where these clearly "disadvantageous" closed models are replaced with superior open models, potentially saving $104-146 million annually on the OpenRouter platform alone. Extending this to the entire inference market, three independent estimation methods show potential but unrealized savings of $20.1 billion to $48.3 billion per year, with the preferred estimate being $24.8 billion.

Trust and switching costs in the AI adoption process

It's worth noting that Nagle and Yue did not argue that user behavior was irrational. Their research shows that factors such as switching costs, brand trust, security concerns, and the possibility that standardized benchmarks may not reflect actual performance differences collectively explain users' preference for closed models. The paper points out that the economic impact of these hidden factors is far greater than previously thought.

Surman cited a study commissioned by the U.S. Department of Commerce, pointing out that there is no significant difference in marginal risk between open-weighted models and closed-weighted models. "Both can be compromised," he said. "The harm caused by misinformation and deepfakes depends on the identity and methods of the deployers. At Mozilla, we participate in the ROOST (Robust Open Online Safety Tools) Foundation project, which develops AI-based trust and security software for platforms to detect and remove harmful content."

Surman further explained: "The business model of open-source AI includes service fees, technical support fees, and allowing users of the same software to share resources. Open-source technology has always been known for its auditability and transparency—anyone can view the code, and when problems arise, many developers can work together to fix them. In contrast, proprietary systems in the AI field are largely unregulated; the key is who you choose to trust."

Surman points out that the technology industry has long been built on open source. He cites cloud service providers such as Amazon Web Services (AWS), Microsoft Azure, and Google Cloud Platform, as well as companies like Meta and X, all of which run on open source software such as Linux.

Regarding the ongoing debate surrounding licensing agreements, Surman explained Mozilla's position: "Permissive licensing remains the right choice. Open source leverages copyright law through licensing agreements, granting the right to modify and share the results. To achieve responsible AI applications, we need safety barriers, business processes, and national laws, not to change open source licenses."

As the global order is being restructured, countries are actively exploring ways to promote broad participation and create new space for open-source AI. Facing the AI wave hailed as a new era of digital technology, Surman reviews the evolution of the computer age, the personal computer age, the internet age, and points out that open source has always been a crucial force against centralization.

"If open-source AI succeeds, it means that more people can participate in shaping this new era of society, economy, and technology," Surman said. "Five years from now, I expect to see AI sources become more diverse, with more people becoming creators rather than just consumers."