# Introduction
For the last two years, the assumption in agentic AI was simple: the bigger the model, the better the agent. Bigger context windows, more parameters, sharper reasoning. What’s not to like? NVIDIA’s own research team spent 2025 quietly building the case against that assumption, and the argument that came out of it has reshaped how a lot of production agents are actually built in 2026. Most of what an agent does day to day isn’t broad, creative, or novel. It’s a small number of specialized tasks performed repetitively with little variation, and a model trained to be a generalist is overkill for work that’s fundamentally narrow. That’s the opening for small language models (SLMs), and they’ve gone from a footnote to a real architectural decision in agent design.
This article looks at five concrete ways SLMs are showing up inside next-generation agents right now, from the research backing them to the tools and numbers worth knowing if you’re deciding whether your next agent needs a frontier model at all.
# 1. Handling the Repetitive Work Frontier Models Were Never Built For
The foundational case for SLMs in agents comes from a single, widely discussed paper out of NVIDIA Research: Small Language Models are the Future of Agentic AI. The authors argue that large language models are valued for general conversation, but agentic systems mostly call on language models to perform a small number of specialized tasks repetitively, such as parsing a command, selecting a tool, or returning a result in a fixed JSON shape. That’s a completely different job than holding an open-ended conversation, and it doesn’t require a model trained to do everything.
The paper’s central claim is direct: SLMs are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations in agentic systems, and are therefore the future of agentic AI. What makes this more than an opinion piece is the reasoning behind it. Agents value reliability over creativity, and a small model fine-tuned to always follow a fixed output format and field order is often more dependable for that one job than a large general-purpose model asked to do the same thing on the fly. Large models still earn their place for genuinely novel or open-ended reasoning. They’ve just stopped being the default for everything in between.
# 2. Running Directly on the Device, No Cloud Round Trip Required
One of the most practical shifts SLMs have unlocked is moving the model itself off a remote server and onto whatever hardware the agent is already running on, such as a phone, a laptop, a piece of industrial equipment. Sending a request to a data center takes hundreds of milliseconds, while edge inference happens in tens, and for an agent meant to feel responsive in the moment, that gap is the difference between something that feels instant and something that feels like it’s thinking too hard.
The hardware caught up faster than most people expected. The Apple A19 Pro’s neural accelerators give the iPhone 17 Pro enough combined AI throughput to run 8-billion-parameter models at over 20 tokens per second — fast enough for real-time conversation — and Apple’s M5 Max can handle models up to 30B parameters with acceptable latency. Quantization is a big part of why this works on consumer hardware at all. A Phi-4-Mini model compressed to 4-bit precision occupies roughly 1.2 GB of memory instead of 7.6 GB in full precision, while keeping over 95% of its benchmark performance — small enough to fit comfortably on a phone with 8 GB of RAM.
Tools like Ollama for local serving and Microsoft’s Phi model family have become common starting points for developers building this kind of on-device agent behavior, particularly for use cases where the agent needs to keep working even when a network connection isn’t guaranteed.
# 3. Getting Fine-Tuned Into Tool-Calling Specialists
A generic small model, straight out of the box, is genuinely bad at tool calling. It hallucinates function names, gets the parameters wrong, and breaks the expected output format more often than you’d like. The fix isn’t a bigger model — it’s a more focused one. Fine-tuning a small model on a specific tool schema produces accuracy above 90% at effectively zero per-query cost, because the model stops trying to be a generalist and starts being very good at exactly one narrow job.
The research backing this is striking. A fine-tuned SLM achieved a 77.55% pass rate on the ToolBench evaluation, well ahead of baseline approaches using much larger models prompted with chain-of-thought reasoning. You don’t need a huge training run to get there either. In practice, 1,000 to 5,000 high-quality examples per tool are typically enough to reach 95%-plus accuracy on a well-defined schema, which is a realistic amount of data for a small team to produce in-house.
If you want a closer look at which specific models are leading on this right now, KDnuggets recently rounded up five small, open-weight models built specifically for agentic tool calling, spanning a few billion parameters each and built to run without a data center behind them.
# 4. Powering Heterogeneous Systems Where Big and Small Models Split the Work
The most architecturally interesting use of SLMs isn’t replacing large models outright; it’s pairing them. The pattern that’s become standard in 2026 puts a high-reasoning frontier model in the role of planner, handling strategy and ambiguity resolution, while domain-specific small models act as the workers, each fine-tuned for one atomic task like parsing, classification, or summarization. Some call this the executive-worker architecture; others call it heterogeneous model routing. Either way, the idea is to spend the expensive reasoning where it’s actually needed and let cheaper models handle the volume.
The cost difference this produces is hard to ignore. A frontier model priced around \$15 per million tokens handling 30% of tasks, paired with a small model around \$0.15 per million tokens handling the remaining 70%, costs roughly 10 times less than routing everything through the frontier model alone. The pattern holds up in controlled research too. One study comparing a homogeneous setup of all 7B-parameter agents against a heterogeneous setup, where smaller 3B models handled lower-level work while a 7B model stayed on as the verifier, found the heterogeneous system kept performance nearly identical to the all-7B baseline while cutting latency by 31.6% and total API cost by 41.8%. NVIDIA has packaged tooling for building this kind of system into NeMo, aimed at teams that want to mix fine-tuned SLMs for routine work with occasional calls to a larger model for the genuinely hard cases.
# 5. Keeping Sensitive Data on the Device Instead of Sending It Anywhere
The last shift is less about speed or cost and more about where data is allowed to go in the first place. An agent that runs entirely on local hardware never has to send a user’s conversation, documents, or behavior to a third-party API to get a response, which matters a great deal once you’re working with healthcare records, financial information, or anything covered by strict compliance rules.
For healthcare or industrial security use cases specifically, data often cannot leave the local network at all, which rules out cloud-hosted frontier models as an option, regardless of how good they are. Small models sidestep that constraint entirely by running where the data already lives. Edge deployment on something like Apple Silicon or a Qualcomm chip costs only the device hardware itself, with hosting for a private small model serving 10,000 daily queries typically running \$500 to \$2,000 a month, compared to \$5,000 to $50,000 a month for the equivalent volume through a large model API.
This is also the only realistic option for fully air-gapped environments — places with no internet connection at all by design — where a cloud-dependent agent simply can’t function regardless of budget. For agents built for regulated industries or offline-first products, this isn’t a nice-to-have. It’s the entire reason the agent can exist in that environment at all.
# Wrapping Up
None of this means frontier models are on their way out. Genuinely novel reasoning, long open-ended context, and tasks nobody’s seen before still belong to the big models, and that’s not changing anytime soon. What’s changed is the assumption that every single call an agent makes needs that level of horsepower. Most of an agent’s actual work — the parsing, the routing, the formatting, the tool calls — turns out to be narrow enough that a small, fine-tuned model handles it just as well, often faster, and at a fraction of the cost.
The agents that scale well in 2026 aren’t the ones built on the single biggest model available. They’re the ones built with the right-sized model for each piece of the job — frontier intelligence where it’s earned and small, specialized models everywhere else.
Shittu Olumide is a software engineer and technical writer passionate about leveraging cutting-edge technologies to craft compelling narratives, with a keen eye for detail and a knack for simplifying complex concepts. You can also find Shittu on Twitter.
