Research
We think intelligence does not have to be one enormous model in a distant datacentre. It can be a society of small specialists, each expert at one thing, orchestrated and composed on the device in front of you. Our research runs from the single specialist, to the colony that coordinates them, to the compression that makes each one small, to the physics that could run them for a fraction of the power.
Research Area 01
Fern is the unit of our approach: a small model trained for one job. It runs on the device in front of you, a phone, a watch, an embedded chip, fully offline, and it answers in well under a second. It does not write a paragraph about what to do. It returns the action itself, the exact function call, intent turned into a command your software can run.
New Now shipping as Fernfly: Fern's research, productized for developers. →At a glance
Research Area 02
No single ant is smart. The colony is.
Our bet is that intelligence can come from a society of small specialists rather than one large generalist. Many Ferns, each expert at one narrow thing and limited on its own, talk to each other, route and hand off work, and compose into something that behaves like one capable mind, at a fraction of the energy and cost, and all of it running locally.
The name is not a coincidence. Antelligent is ant plus intelligent, and our logo is a colony. An ant colony is the canonical example of emergent intelligence from many simple agents: no single ant is smart, the colony is. The Colony is both what we call this research and the metaphor that drives it.
This is the research that sits underneath the Everyday Series, our product for orchestration and governance. The hard questions get worked out here. The product ships there.
Open research questions
Which specialist should handle a task, and when to escalate to a larger model.
Passing state between specialists without a giant shared context, the thing that makes big-model orchestration so token-expensive.
Whether a society of small models can match or beat one large model at equal cost, and where that crossover falls.
Doing this coordination at the edge rather than in a datacentre.
Active research directions
Pushing compression further on next-generation model families without quality regression.
Formal quality guarantees for regulated domains such as medical and defence.
Extending our compression pipeline to protein, diffusion, and vision-language models.
Building compressibility into the model from the start rather than applying it post-hoc.
Research Area 03
Our compression technology already produces models with substantially fewer parameters, with quality measured and reported. But we're not done. Our research keeps advancing: broader model coverage, and compression techniques that are provably safe for regulated environments.
Active directions include architecture-aware pruning for next-generation model families, provable quality bounds for safety-critical applications, and compression-native training pipelines that build compressibility into the model from the start.
These improvements feed directly into our products: better compression means smaller on-device footprints, lower enterprise serving costs, and higher-quality compressed models in our open-source releases.
Research Area 04
The furthest-out of our bets is about the physics of computing itself. Instead of a GPU calculating an answer step by step, let a set of coupled physical oscillators settle into one. You pose the problem as a physical system, let it relax toward its lowest-energy state, and read the answer off where it lands.
The interesting behaviour sits at the edge of chaos, the narrow band between rigid order and noise where a physical system is most expressive. Pair that with compute-in-memory, where storage and arithmetic are the same physical device rather than numbers shuttled back and forth, and the aim is orders of magnitude less power than a conventional chip.
We are honest about where this stands. This is early research, a bet on a direction rather than a product, and one path among several we are watching. We call early research early, and we publish the numbers that go against us alongside the ones that do not.
What we are exploring
Coupled oscillators relax into an answer instead of a processor computing one step by step.
The expressive band between rigid order and noise, where a physical system carries the most information.
Memory and arithmetic in the same device, so a value is not carried across the chip to be worked on.
The reason to try at all: the target is a physical process that costs a small fraction of what silicon does today.
Collaborate
We're looking for hardware partners, academic collaborators, and engineers who want to work on the foundations of sustainable AI, from the single specialist to the colony to the silicon underneath.
