Almost every business machine on earth has a capable processor and some spare memory. Almost none has a datacentre GPU. We think that is the interesting constraint, and it is the one we are building for.
There is an assumption buried in almost every AI product built in the last three years: that the model runs somewhere else. You type, the request crosses the internet to a rack of specialised accelerators, and the answer comes back. It is a good arrangement and it produced most of what is useful today.
It also has consequences nobody chose deliberately. Cost scales with use, forever. Capability depends on a connection. And the economics of the entire industry rest on access to a class of hardware that is scarce, expensive, and allocated by a small number of suppliers.
So we started asking a different question. Not “how large can a model be”, but how much useful work can be done on the hardware already sitting on people’s desks?
Inference in someone else’s datacentre is a per-use cost that never ends. Every summary, every draft, every background job has a price, and the better the tool the more of them you run. Work that happens on a machine you already own has a different shape: the marginal cost of one more task is close to nothing.
For an individual that is a convenience. For a business running a worker across thousands of documents a month, it is the difference between a feature and a line item somebody eventually cancels.
A great deal of Australian work happens where the internet is poor or absent. Sites, mines, vessels, remote clinics, aircraft, basements. Any capability that needs a round trip to a datacentre simply is not available in those places, and the people doing that work are used to being last in line for new software.
Network round trips have a floor set by physics and routing, not by budget. For anything genuinely interactive — a suggestion while you type, a check that runs as you edit — local execution is not merely cheaper, it is a different experience.
The strongest privacy guarantee is structural: data that never leaves the machine cannot be intercepted, retained or compelled in transit, because it was never in transit. That is a property of design rather than of promises, and it is the version that professionals with confidentiality obligations actually find persuasive.
If this were easy it would already be done. Two constraints do most of the damage.
The bottleneck is memory, not arithmetic. Generating each token means reading a large set of weights out of memory. On ordinary hardware, throughput is limited by how fast memory can be read long before the processor runs out of maths to do. This is why “my CPU has plenty of cores” does not translate into speed the way people expect it to.
Compression helps, but it is not free, and it does not transfer. Quantisation — storing weights at lower precision — is the main lever, and it works. But a study of language model inference at the edge found that while “quantization helps mitigate memory overhead, it does not fully eliminate resource bottlenecks, especially for larger models”. And a benchmark of quantisation applied specifically to small models found that “direct transfer of LLM-optimized techniques leads to suboptimal results due to SLMs’ unique architectural characteristics”, citing “fundamental disparities between SLMs and LLMs in quantization sensitivity”.
That second finding is the interesting one, and it is why this is a research problem rather than an engineering chore. You cannot take the recipes that made large models efficient, apply them to a small model, and expect them to hold. The small case needs its own answers, and comparatively few people are looking for them.
Oppermind already develops and runs its own language models. They serve parts of the Oppermind workspace, and we deploy custom versions of them for corporate partners who want a system that runs on their infrastructure, tuned to their work. That part is not aspiration — it is what we already do.
The public-facing extension of that work is what this article is about: a model you can install on the computer you already own, without a datacentre GPU. Not a reduced version of something built for accelerators, but a design that treats a normal processor and a few gigabytes of memory as the environment it has to live in.
We are not chasing the largest general-purpose model on the leaderboard. That is a different sport with different economics. What we are willing to be judged on is narrower and, we think, more useful: whether real work can be done dependably on machines people already have, at a cost that does not climb with every use.
Efficiency research has a structural credibility problem: it is easy to report a favourable number and very hard for anyone outside to check it. Benchmarks can be chosen after the results are in. Test hardware can be unusually well suited. The comparison can be against something nobody would have used anyway.
The honest response is to publish enough that somebody can disagree with you using evidence. That is our intention with this work — and we are deliberately calling it an intention rather than a commitment with a date, because research schedules that get announced tend to produce worse research.
If it works, an open result is worth more to us than a closed one: we sell a workspace, not model weights, and a healthy ecosystem of small efficient models makes the product better no matter who builds them. If it does not work, that is worth publishing too. The field has enough quiet failures.
If you use the Oppermind workspace, some of what you already run is powered by our own models — that has been true for a while and there is nothing on this page you need to enable to benefit from it. If you are a corporate partner who needs a system that runs on your own hardware, we build custom deployments through Corporate Solutions today. What is not yet available is a downloadable model for your own PC, and no date is being announced here.
If you want the practical companion to this piece — what genuinely does run on an ordinary machine right now, and how to judge whether it is fast enough to be useful — that is can you run an LLM on a normal PC?
And if you simply want the workspace, it is free with no card, then Starter A$9.95, Pro A$29.99 and Pro Plus A$59.99 a month, with unlimited documents, spreadsheets, decks and designs on every plan.
Whatever you’re here to make, make more of it.
Yes. Our own local models power parts of the Oppermind workspace today, and we build custom deployments of them for corporate partners through Corporate Solutions. What is not yet available is a downloadable local model you install on your own PC — that is the direction this article is about, and we will say clearly when it ships.
Because of what is already installed. Almost every business computer on earth has a capable CPU and a few gigabytes of spare memory, and almost none has a datacentre GPU. A model that needs specialised hardware can only run where that hardware is; a model that runs on ordinary machines can run where the work actually happens, including offline.
Mostly memory rather than arithmetic. Generating each token requires reading model weights from memory, so throughput is limited by memory bandwidth long before the processor runs out of maths. Compression helps but does not eliminate the bottleneck, and research indicates that compression techniques tuned for large models transfer poorly to small ones.
That is our intention, and we are stating it as an intention rather than a commitment with a date. Research that cannot be inspected is difficult to trust, particularly research about efficiency, where it is easy to report a favourable number and hard for anyone outside to check it. We would rather publish and be corrected.
Not on their own ground. We are not chasing the largest general-purpose model on the leaderboard. Our own models are aimed at doing useful work reliably, on hardware people already have, at a predictable cost — for the workspace, for our corporate partners, and eventually as something you can install yourself.
Parts of this article describe the direction of an on-device, publicly-installable model that is not yet released; those parts are forward-looking, and plans and timing may change. Nothing on this page announces a public release date, promises a specific future capability, or quotes a performance figure for an Oppermind model. Statements about our existing use of our own models to power the Oppermind workspace and Corporate Solutions deployments are current as at 15 September 2026. Third-party research is cited with its date and is the work of its authors, not ours. Oppermind platform prices are current as at 15 September 2026, are in Australian dollars, and are subject to the plan terms at checkout.
A full suite of AI tools and editors, with the autonomous worker included in every plan.