A 32MB speech model is the more interesting number

Buried under the funding announcements this week was a demo from a company called Underdog: a speech model that runs at 32 megabytes.

Thirty-two megabytes. For scale, that is smaller than a single photo from a modern phone camera, and about a sixtieth of the smallest model most vendors would have called "small" two years ago. It does not need a data centre. It does not need a connection. It fits in the app you already have installed.

Nobody is going to raise $11 billion on the back of that. Which is precisely why it deserves more attention than the round that did.

A worker on a factory floor holding a phone, machinery out of focus behind — The useful question is not how smart the model is. It is where it runs.

The assumption that has been quietly doing the work

Every AI tool sold to an export team in the last three years has rested on one unstated assumption: that the model runs somewhere else, and your data goes to it.

That assumption was reasonable while models were huge. It is becoming a choice rather than a constraint, and the choice has consequences that procurement teams rarely price in:

Data leaves the building. A supplier's spec sheet, a factory's price list, a buyer's email — all of it crosses a border to a server you do not control. For a trading company with European customers, that is a compliance question dressed as a convenience.
The tool stops when the signal stops. Walk into a factory in Dongguan with two bars of 4G and watch a cloud transcription tool fail. The work does not pause because the model is unreachable.
Latency is a tax on every interaction. Two seconds per utterance is fine for a meeting summary and unbearable for a live conversation with a supplier across a table.
You pay per token for something that could be free at the margin. A model that runs locally costs electricity, not API calls.

None of these are new arguments. What is new is that the size threshold has dropped far enough that "runs on the device" is now an option for real work rather than a demo trick.

What a small model is actually good at

A 32MB speech model is not going to write your buyer emails. It will not reason about a tariff schedule. Be clear about the trade: you lose breadth and world knowledge, and you gain privacy, offline operation, latency measured in tens of milliseconds, and a marginal cost near zero.

For export operations, that trade is often the right one, because the tasks are narrow and repetitive:

transcribing a supplier call in a noisy workshop
dictating inspection notes while your hands are full
translating a short exchange on the floor, without uploading the conversation
turning a spoken spec into structured fields

Each of those is a bounded task with a fixed vocabulary. A small model tuned for it does not need to know what a tariff is. It needs to hear "two thousand pieces, FOB Shenzhen" correctly, every time, in a room with a lathe running.

The best model for a factory is not the smartest one available. It is the one that is still working when the wifi drops.

What to ask before you buy the next "AI feature"

We have started asking three questions that would have sounded eccentric a year ago:

Where does inference run? If the answer is "our cloud" for every task, ask which tasks could run on-device and why they do not. The answer is usually "we did not build it that way", which is a fair answer but a negotiable one.

What happens offline? A tool that degrades gracefully — queueing, local draft, later sync — is a different product from one that shows a spinner. Test it in airplane mode before you sign.

What is the model size, and does it change? Vendors rarely publish this, and it matters: a 30MB model can run on the phone in your pocket, a 3GB one cannot. If they will not say, assume the worst and price accordingly.

The same logic applies to the laptop-based tools your team already uses. A translation model that runs locally on a laptop handles a 200-page spec sheet without sending it anywhere, and for a company whose differentiator is knowing its suppliers, not uploading them, that is worth more than a marginal quality gain from a bigger model.

A workshop bench with a phone, a digital caliper and a small circuit board — Small enough to live in the tool you already carry. That is the whole argument.

The boring conclusion, which happens to be right

Big models will keep getting better and the funding will keep flowing to whoever trains them. That story is real and it is not going to stop being the headline.

But the practical shape of AI in export work over the next two years is not one giant model answering everything. It is a dozen small ones, each doing one narrow job, many of them running on hardware you already own, most of them invisible.

The companies that get value from this will not be the ones with the best model access. They will be the ones who noticed that the constraint moved — from "can it do the task" to "can it do the task here, on this device, without a connection, without sending the document to a stranger".

Thirty-two megabytes is a small number. It is also the first one this week that changed what a sourcing desk can actually do.

We test field tools the same way we test everything else: on a real task, in a real place, with the wifi off. Tools that pass get a note in the listing. Tools that only work in a demo do not.