Everyone’s watching LLMs dominate the spotlight. But zoom in a little closer, and something else is brewing. Small Language Models (SLMs) are lighter, faster, and more accessible. For agentic AI and decentralized systems, they might be a better fit. Here’s why 🧵

2/ LLMs are just "large". Trillions of parameters, powerful generalists, and expensive to run. They work great for broad, open-ended tasks. But they’re centralized, opaque, and hard to customize. SLMs are compact, transparent, and flexible. You can fine-tune and run them on your terms.

3/ SLMs shine in real-world settings: They’re efficient, quick to respond, and don’t need heavy infrastructure. Perfect for edge devices and privacy-sensitive use cases. With tools like distillation, pruning, and test-time reasoning, they deliver serious performance at a fraction of the cost.

4/ New research proves it. The Alan Turing Institute ran a 3B parameter model on a laptop. With smart tuning, it nearly matched frontier models on health reasoning tasks. This wave is growing: Phi, Nemotron-H, Qwen3, Mu, SmolLLM, all pushing SLMs into the mainstream.

5/ More details of this research:

6/ We believe in building AI that’s open, local, and decentralized. SLMs make that possible. They’re not just lightweight alternatives, they’re the groundwork for scalable, modular agentic systems. Small isn’t a compromise. Small is a powerful design choice.

7/ Read our full blog on this topic 👇