Most of the AI tooling you actually need day to day will run on a model small enough to sit on your own hardware, and the industry is only just admitting it.
I review toolkits for a living, which means I spend an unglamorous amount of time watching people pay for capability they never touch. The pattern is consistent. A team wires up the largest available frontier model, routes every request through it, then discovers eight months later that ninety percent of those requests were classification, extraction, summarization, or reformatting. Tasks a much smaller model handles fine. Tasks a much smaller model handles fine and faster and without a per-token meter running.
So the news that Australia is leading a rise in small language models, per the AFR, does not read to me as a curiosity from the other side of the world. It reads as the first mainstream acknowledgment of something that has been true in practice for a while.
Why the timing makes sense
Two other stories landed in the same news cycle and they are more connected than they look.
The first is MIT Technology Review reporting that AI’s recursive self-improvement might not come so quickly after all. That matters for tooling decisions in a very practical way. If you believe a self-improving system is six months out, the rational move is to wait. Buy the biggest thing, hold your architecture loose, assume everything you build gets obsoleted. If that timeline stretches, the calculus flips. Now you are optimizing for the models that exist, on the hardware you have, at the cost you can defend to whoever signs off on your invoices.
The second is Rest of World’s piece on Americans choosing Chinese AI. Strip out the geopolitics for a second and look at what that behavior signals. Users are picking tools on price and availability rather than brand loyalty or benchmark supremacy. That is a commodity market forming in real time. And commodity markets reward the cheapest thing that clears the bar, not the most capable thing on the shelf.
Small models are the cheapest thing that clears the bar for an enormous share of real work.
What the hardware story tells you
AppleInsider’s iPhone 18 rumors mention 12GB or 9GB of RAM and a 2nm A20 chip. I have no idea if those specs hold. Rumors two years out are rumors. But the direction is the interesting part, because RAM is the constraint that determines what runs on a phone versus what has to make a round trip to somebody’s data center.
Every gigabyte added to a phone’s memory budget moves the line on what can run locally. And local changes the whole shape of a product. No network dependency. No latency on the wire. No inference bill that scales with your user count. For a lot of the tools I look at, that is not a marginal improvement, it is the difference between a viable business and a slow bleed.
The part reviewers get wrong
I include myself here. The reviewing culture around AI tools has been organized around ceiling. What is the hardest thing this model can do. Which one wins on the tricky reasoning prompt. It makes for readable comparisons and it is close to useless for procurement.
What actually determines whether a toolkit works in production:
- Does it hold up on your specific, boring, repetitive task at volume
- What does it cost when the request count goes up by 50x
- How fast does it respond when a user is waiting on screen
- Can you run it somewhere you control, if you need to
- What happens when the vendor changes the model underneath you
Small models score well on most of those and badly on ceiling. Which is why they got dismissed for a couple of years and why they are being rediscovered now.
Where I’d still be careful
This is not a blanket recommendation. Small models fall over on genuinely hard multi-step reasoning, and if your product depends on that, no amount of cost savings fixes a wrong answer. The honest architecture is mixed: small models on the volume work, something larger on the narrow slice that needs it, and a routing layer you understand well enough to debug.
The other risk is fragmentation. Top Gear listed 44 new electric cars coming, which is a fine analogy for where model selection is heading. Abundant options, most of them adequate, no obvious way to choose without doing the work yourself. That is a better problem than having no options, but it is still a problem, and it lands on whoever has to evaluate.
My read: pick a small model, test it on your actual workload rather than a benchmark, and only reach for something bigger when you have evidence you need it. Most teams will need it less often than they expect.
đź•’ Published: