Small language models capable of addressing specific tasks without burning through processing power are likely to surge in popularity. Credit: iQoncept - shutterstock.com In 2024, the surge in generative AI (genAI) pilot projects sparked concerns over high experimentation costs and uncertain benefits. That prompted companies to then shift their focus to delivering business outcomes, enhancing data quality, and developing talent. In 2025, enterprises are expected to prioritize strategy, add business-IT partnerships to assist with genAI projects and move from large language model (LLM) pilots to production instances. And small language models will also likely come into their own, addressing specific tasks without overburdening data center processing and power. Organizations will also adopt new technologies and architectures to better govern data and AI, with a return to predictive AI, according to Forrester Research. Predictive AI uses historical data and techniques such as machine learning and statistics to forecast future events or behaviors, said Forrester analyst Jayesh Chaurasia. GenAI, on the other hand, creates new content — such as images, text, videos, or synthetic data — leveraging deep learning methods such as generative adversarial networks (GANs). Chaurasia predicts the AI pendulum will swing back to predictive AI for over 50% of use cases. LLMs are, of course, central to genAI, helping enterprises tackle complex tasks and improve operations. Forrester reported that 55% of US genAI decision-makers with a strategy use LLMs embedded in applications, while 33% purchase domain-specific genAI apps. Meanwhile, SLMs are quickly gaining attention. The rise of small and mid-sized language models should enable customers to better meet the trade-offs on accuracy, speed and costs, said Arun Chandrasekaran, a distinguished vice president analyst with Gartner Research, noting that “Most organizations are still struggling to realize business value from their genAI investment.” Gartner In the coming year, SLM integration could surge by as much as 60%, according to a Forrester report. As nearly eight-in-10 IT decision makers report software costs rising over the past year, they’re looking to SLMs because they’re more cost-effective and offer better accuracy, relevance, and trustworthiness by training on specific domains. They’re also easier to integrate and excel in specialized industries such as finance, healthcare, and legal services. By 2025, 750 million apps are expected to use LLMs, underscoring the genAI market’s rapid growth. Forrester predicts the market will grow in value from $1.59 billion in 2023 to $259.8 billion by 2030,. Even with that growth, many AI experts argue that LLMs may be excessive for automating workflows and repetitive tasks, both in terms of performance and environmental impact. A Cornell University study found that training OpenAI’s GPT-3 LLM consumed 500 metric tons of carbon, the equivalent of 1.1 million pounds. As enterprises face challenges meeting expectations, gen AI investments in 2025 will likely shift toward proven predictive AI applications like maintenance, personalization, supply chain optimization, and demand forecasting. Forward-thinking organizations will also recognize the synergy between predictive and generative AI, using predictions to enhance generative outputs. That approach is expected to boost the share of combined use cases from 28% today to 35%, according to Forrester. What are small language models? In the simplest of terms, an SLM is a lightweight genAI model. The “small” in this context refers to the size of the model’s neural network, the number of parameters, and the volume of data on which it is trained, according to Rosemary Thomas, a senior technical researcher in the AI lab at Version 1, a management consulting and software development firm. SLMs use fewer computational resources, enabling on-premises or private cloud deployment, which natively enhances privacy and security. While some SLM implementations can require substantial compute and memory resources, several models can have more than 5 billion parameters and run on a single GPU, Thomas said. Gartner Research defines SLMs differently, as language models with 10 billion parameters or less. Compared to LLMs, they are two to three orders of magnitude (around 100-1,000x) smaller, making them significantly more cost-efficient to use or customize. SLMs include Google Gemini Nano, Microsoft’s Orca-2–7b and Orca-2–13b, Meta’s Llama-2–13b, and others, Thomas noted in a recent post, arguing that SLM growth is being driven by the need for more efficient models and the speed at which they can be trained and set up. Gartner “SLMs have gained popularity due to practical considerations such as computational resources, training time, and specific application requirements,” Thomas said. “Over the past couple of years, SLMs have become increasingly relevant, especially in scenarios where sustainability and efficiency are crucial.” SLMs enable most organizations to achieve task specialization, improving the accuracy, robustness, and reliability of genAI solutions, according to Gartner. And because deployment costs, data privacy, and risk mitigation are key challenges when using genAI, SLMs offer a cost-effective and energy-efficient alternative to LLMs for most organizations, Gartner said. Three out of four (75%) of IT-decision makers believe SLMs outperform LLMs in speed, cost, accuracy and ROI, according to a Harris Poll of more than 500 users commissioned by the start-up Hyperscience. “Data is the lifeblood of any AI initiative, and the success of these projects hinges on the quality of the data that feeds the models,” said Andrew Joiner, CEO of Hyperscience, which develops AI-based office work automation tools. “Alarmingly, three out of five decision makers report their lack of understanding of their own data inhibits their ability to utilize genAI to its maximum potential. The true potential…lies in adopting tailored SLMs, which can transform document processing and enhance operational efficiency.” Gartner recommends that organizations customize SLMs to specific needs for better accuracy, robustness, and efficiency. “Task specialization improves alignment, while embedding static organizational knowledge reduces costs. Dynamic information can still be provided as needed, making this hybrid approach both effective and efficient,” the research firm said. In highly regulated industries, such as financial services, healthcare and pharmaceuticals, the future of LLMs is definitely small, according to Emmanuel Walckenaer, CEO of Yseop, a vendor that offers pre-trained genAI models for the BioPharma industry. Smaller, more specialized models will reduce wasted time and energy spent on building large models that aren’t needed for current tasks, according to Yseop. Agentic AI holds promise, but it’s not yet mature In the year ahead, there is likely to be a rise in domain-specific AI agents, “although it is unclear how many of these agents can live up to the lofty expectations,” according to Gartner’s Chandrasekaran. While Agentic AI architectures are a top emerging technology, they’re still two years away from reaching the lofty automation expected of them, according to Forrester. While companies are eager to push genAI into complex tasks through AI agents, the technology remains challenging to develop because it mostly relies on synergies between multiple models, customization through retrieval augmented generation (RAG), and specialized expertise. “Aligning these components for specific outcomes is an unresolved hurdle, leaving developers frustrated,” Forrester said in its report. A recent Capital One survey of 4,000 business leaders and technical practitioners across industries found that while 87% believe their data ecosystem is ready for AI at scale, 70% of technologists spend hours daily fixing data issues. Still, Capital One’s survey revealed strong optimism among business leaders about their companies’ AI readiness. Notably, 87% believe they have a modern data ecosystem for scaling AI solutions, 84% report having centralized tools and processes for data management, 82% are confident in their data strategy for AI adoption, and 78% feel prepared to manage the increasing volume and complexity of AI-driven data. And yet, 75% of enterprises attempting to build AI agents in-house next year are expected to fail, opting instead for consulting services or pre-integrated agents from existing software vendors. To address the mismatch between AI data preparedness and real-world complexities in 2025, 30% of enterprise CIOs will integrate Chief Data Officers (CDOs) into their IT teams as they lead AI initiatives, according to Forrester Research. CEOs will rely on CIOs to bridge the gap between technical and business expertise, recognizing that successful AI requires both solid data foundations and effective stakeholder collaboration. Forrester’s 2024 survey also showed that 39% of senior data leaders report to CIOs, with a similar 37% reporting to CEOs — and that trend is growing. To drive AI success, CIOs and CEOs must elevate CDOs beyond being mere liaisons, positioning them as key leaders in AI strategy, change management, and delivering ROI. A growing interest in multi-modality — and upskilling Emerging use cases for multi-modality, particularly image and speech as modalities in both genAI inputs and outputs, will also see more adoption in 2025. Multimodal learning, a subfield of AI, enhances machine learning by training models on diverse data types, including text, images, videos, and audio. The approach enables models to identify patterns and correlations between text and associated sensory data. By integrating multiple data types, multimodal AI expands the capabilities of intelligent systems. These models can process various input types and generate diverse outputs. For example, GPT-4, the foundation of ChatGPT, accepts both text and image inputs to produce text outputs, while OpenAI’s Sora model generates videos from text. Other examples include medical imaging, patient history, and lab results that can be integrated to enhance pateitn diagnosis and treatment. In financial services, multimodal AI can analyze customer phone queries to assist contact center employees in resolving issues. And in the automotive industry inputs from cameras, GPS, and LiDAR can be integrated by AI to enhance autonomous driving, emergency response, and navigation for companies, such as Tesla, Waymo and Li Auto. AI leaders have also realized they need to prioritize business outcomes, clean their data houses, and start building AI talent. The latter is especially important given the growing gap between enterprise genAI needs and the workers with skills to meet those needs. “In the year ahead, you’ll need to put your nose to the grindstone to develop an effective AI strategy and implementation plan,” Forrester said in its report. “In 2025, organizational success will depend on strong leadership, strategic refinement, and recalibration of enterprise data and AI initiatives commensurate with AI aspirations.” SUBSCRIBE TO OUR NEWSLETTER From our editors straight to your inbox Get started by entering your email address below. Please enter a valid email address Subscribe