ELUXR

Small Action Models Are the Future of AI Agents

2025 is the year of agents, & the key capability of agents is calling tools. Imagine a scenario where an AI can streamline processes for businesses by executing tasks with the mere command of a user. For example, using Claude Code, an AI could sift through a newsletter, find all the links to startups, and verify their existence in a CRM with a single instruction. This kind of automation is not just futuristic; it’s already becoming a reality.

However, there exists a challenge: relying on large foundation models for seemingly simple selection tasks can be expensive, rate-limited, and often overpowered. The question remains: what is the ideal way to build an agentic system that effectively utilizes tool calling?

The Answer: Small Action Models

The answer lies in small action models. A compelling paper by NVIDIA suggests that “Small language models (SLMs) are sufficiently powerful and inherently more suitable for many invocations in agentic systems.” Testing local models has confirmed that this approach leads to cost reductions without sacrificing functionality. For instance, one could start with a Qwen3:30b parameter model. While functional, it is notably slower due to its size, operating with only 3 billion of its 30 billion parameters at any given time.

NVIDIA recommends leveraging the Salesforce xLAM model – a different architecture specifically designed for efficient tool calls. A recent test comparing performance among various models in a real-world setting demonstrated compelling results when listing Asana tasks:

• xLAM: 100% success rate (25/25) with an average total time of 2.61 seconds.
• Qwen: 92% success rate (23/25) with an average total time of 9.82 seconds.

The xLAM model completed tasks in a fraction of the time taken by Qwen, illustrating substantial gains in speed and reliability.

Trade-offs and Considerations

While the performance difference is striking, it surfaces a critical trade-off: how much intelligence should reside in the model versus in the tools themselves? With larger models like Qwen, tools can afford to be simpler because the AI can compensate for any limitations, using brute-force reasoning to overcome poorly designed interfaces. In contrast, smaller models necessitate more robust tools and precise selection logic.

This might appear as a limitation, but it is actually a beneficial feature. By requiring better system design, small action models can help eliminate the compounding error rates that occur with larger models initially chained to several tools. Errors in such a setup tend to accumulate exponentially, complicating matters further.

In this light, small action models champion efficiency. They maintain the advantages of LLMs while amalgamating specialized models to streamline system design, leading to faster, more predictable outputs.

Business Benefits of Small Action Models

Implementing small action models can significantly benefit businesses by optimizing processes and reducing costs associated with AI operations. Below are specific advantages:

1. Cost Efficiency: Smaller models require less computational power, leading to lower operational costs.
2. Increased Speed: Faster task execution translates directly to improved productivity.
3. Higher Success Rates: More precise tool calls result in reduced error rates and greater reliability.

Examples of ROI

Consider a company that automates its report generation process:

• Implementing small action models can cut report generation time from hours to minutes.
• If such a process saves 3 hours per day for 10 employees, the company could see an annual ROI of over $100,000, factoring in labor costs.
• Additionally, reducing errors in reports can lead to better decision-making and increased profitability.

Action Steps for Businesses

To harness these benefits, businesses should:

1. Assess current AI tools and identify areas where small action models could replace larger models.
2. Invest in training for employees on the new systems to maximize usability and efficiency.
3. Conduct pilot tests with small action models to evaluate their effectiveness in specific tasks.

Conclusion

In conclusion, the shift toward small action models in AI agents presents an unprecedented opportunity for businesses looking to enhance efficiency, reduce costs, and improve overall performance. As we embark on this journey into the future of AI agents, consider evaluating your current systems and exploring how small action models can reshape your operations. Schedule a consultation with our team today to discuss how we can help you implement these innovative solutions.

Source