2025: The Rise of Small Action Models – The Future of AI Agents Awaits!
Small Action Models Are the Future of AI Agents
2025 is poised to be the year of AI agents, with a key capability being their ability to call tools. Imagine this scenario: you instruct an AI to sift through a newsletter, find links to startups, and verify their existence in your CRM—all with a single command. This task might involve two or three different tools being employed in a seamless manner.
However, a significant challenge arises: using large foundation models for this specific task can be expensive, often rate-limited, and simply overpowered for a selection task. This raises an essential question: what is the most effective way to build an agentic system with tool calling capabilities?
The Rise of Small Action Models
The answer appears to lie in small action models. A compelling paper by NVIDIA posits that “Small language models (SLMs) are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations in agentic systems.” To explore this further, I began testing various local models to assess their cost-effectiveness.
Initially, I experimented with a Qwen3:30b parameter model. While functional, it proved to be relatively slow due to its size—even though only 3 billion of its 30 billion parameters were active at any one time. Following NVIDIA’s recommendation, I shifted my focus to the Salesforce xLAM model, designed specifically for tool selection.
Performance Testing
To evaluate effectiveness, I conducted a test where each model called a tool to list my Asana tasks. Here are the results:
- Model: xLAM
- Success Rate: 100% (25/25)
- Average Response Time: 1.48s
- Average Tool Time: 1.14s
- Average Total Time: 2.61s ± 0.47s
- Model: Qwen
- Success Rate: 92% (23/25)
- Average Response Time: 8.75s
- Average Tool Time: 1.07s
- Average Total Time: 9.82s ± 1.53s
The results from this experiment were striking. The xLAM completed tasks in an average of 2.61 seconds with 100% success, while Qwen required 9.82 seconds with only 92% success—almost four times as long. This emphasizes the speed gains offered by SLMs.
Intelligence Distribution and System Design
However, there’s a trade-off regarding how much intelligence should reside in the model versus in the tools themselves. Larger models, such as Qwen, provide a safety net; their robustness allows for simpler tools since the models can compensate for poorly designed interfaces. They utilize brute-force reasoning to work around these limitations.
In contrast, smaller models necessitate better-designed tools and precise selection logic since they have less capacity to recover from errors. This may initially seem like a drawback, but it’s actually a beneficial feature. By compelling better system design and minimizing error propagation, small action models eliminate the compounding error rates often seen with chained calls in larger models.
This approach enables a fusion of the best attributes of large language models and specialized models, resulting in a more efficient, faster, and predictable system architecture.
Business Implications and Implementation
- Benefits to Business: Implementing small action models can lead to significant cost savings, improved task completion rates, and enhanced system reliability. Businesses can achieve increased efficiency and productivity.
- Average ROI Examples: By transitioning to SLMs, companies could reduce operational costs by up to 40%, while improving output by 25%, potentially yielding ROI that exceeds several times the initial investment.
- Actions for Implementation:
- Conduct feasibility studies to identify suitable tasks for small action models.
- Invest in training and development to ensure teams understand how to leverage SLMs effectively.
- Gradually integrate small action models into existing workflows, starting with less complex, high-volume tasks.
Conclusion
In summary, small action models are redefining the landscape of AI agents, enabling businesses to execute tasks more efficiently and reliably. As companies contemplate their future strategies, adopting these models can lead to significant improvements in productivity and cost-management. To explore how small action models can enhance your business operations, consider scheduling a consultation with our team today.
