From LLM-Centric Approaches…
The future of Agentic AI will be Open Source, Data-Driven & Small Language Models
…to an open source, data-driven small language model world…
A while back NVIDIA released a study stating that Small Language Models (SLMs) are the future Agentic AI…what I find interesting is the excitement it caused in the market.
I’ve always been fascinated by how open-source communities drive AI innovation, especially when it intersects with efficient, agentic systems.
A recent arXiv paper, Is Open Source the Future of AI? A Data-Driven Approach, looks at this with hard data from Hugging Face and GitHub.
It echoes NVIDIA’s push for Small Language Models (SLMs) in agentic AI, showing how open-source is making smaller, smarter models a reality.
In Short
- Open-source AI is booming, with communities rapidly improving smaller models (under 20B parameters) that rival larger ones in performance.
- Lean models are fine-tuned to be performant for specific tasks and hence multiple SLM’s are orchestrated.
- Skewed contributions from top authors, dominance of architectures like Llama and Mistral, and a shift toward efficient, fine-tuned models.
- This aligns with NVIDIA’s view that SLMs are the future of agentic AI — cheaper, faster, deployable, potentially replacing 40–70% of LLM calls.
- It’s data-driven proof that small, open models could power the next wave of agents.
More on the Study
The paper analyses open-source LLMs using metrics from Hugging Face’s Open LLM Leaderboard and GitHub repos.
It’s a quantitative deep dive into community dynamics, model trends and performance gains — perfect for those of us who crave data over hype.
Considering models,, fine-tuned and chat-focused variants dominate, often built on popular bases like Llama and Mistral.
These get quick community love, leading to iterative improvements.
Smaller is better.
Models under 20B parameters gets 85% of downloads (up to 15B, specifically).
Performance on benchmarks keeps climbing, with compact models closing the gap on big models — proving you don’t need massive scale for solid results.
You don’t need massive scale for solid results.
The study warns of risks like misuse but champions open-source for transparency and trust.
It’s dependent on big tech for base models, suggests proprietary AI might pivot to SaaS built on open foundations.
Small Language Models and NVIDIA’s Take
This resonates deeply with NVIDIA’s recent paper, where they state that Small, rather than large, language models are the future of agentic AI.
They argue SLMs are ideal for AI Agents — those AI systems that plan, reason, and act autonomously — because most tasks are narrow, not needing LLM-level generality.
Why SLMs are well suited in agentic setups, per NVIDIA
Inference efficiency
Faster responses, lower latency for real-time AI Agents.
Fine-tuning agility
Easier to adapt for specific errands.
Edge deployment
Run on devices without cloud dependency.
Parameter utilisation
Better value for money.
In experiments, SLMs were used for 40–70% of LLM calls without major performance drops.
There are tooling gaps and conversion challenges, but NVIDIA proposes a 6-step algorithm to migrate from LLMs to SLMs seamlessly.
## Barriers and the Road Ahead
Adopting this open, small-model future isn’t straightforward.
The study highlights sustainability issues in open-source, over-reliance on a few contributors and calls for better policies on misuse.
NVIDIA echoes this with practical hurdles like SLM fine-tuning tools lagging behind.
Lastly, a data-driven shift to efficient AI is required.
Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and language. Language Models, AI Agents, Agentic Apps, Dev Frameworks & Data-Driven Tools shaping tomorrow.
