Local models are the future. We're hiring the people who'll prove it.
The bet, the hard problems, and the open roles — the honest version.
The thesis in one breath: today, frontier AI lives in a datacenter — you rent it, and it reads everything you feed it. That's a temporary state. The moment a model small enough to run on your own laptop is good enough at the work, the cloud's advantage evaporates and privacy stops being a trade-off. Quenderin is a bet that that moment is close — and that closing the gap is a research problem worth a career.
Why local is where this goes
Renting intelligence from someone else's computer is the current default, not the natural one. When the model runs on your machine, three things that are impossible in the cloud become free:
- Privacy stops being a compromise. Your files, your screen, your intent never leave the device. For anyone in law, medicine, or finance, that's the line between "can use an agent" and "can't" — not a preference.
- You own it. No account, no per-token bill, no rate limit, no outage, no quiet model swap under you. Run it ten thousand times; it costs what running it once did. It works on a plane and in a basement.
- It can actually touch your computer. A local agent can stop dead mid-task, show a plain-text record of everything it did and refused, and undo a whole task later — guarantees a cloud agent structurally can't offer about your own machine.
The one thing standing between here and there is capability — and capability at small scale is an engineering and research problem, not a law of nature. That's the whole job.
The hard problems (what you'd actually work on)
These aren't hypotheticals — they're the open frontier of the codebase today, with every measured result published and every failure written down.
- Small-model agency. Make a 4–8B model a reliable agent: plan a multi-step task, pick the right tool, notice when it's wrong, and recover — on hardware with 8 GB of RAM. This is the gap between "impressive demo" and "does my chores." The mechanism (constrained decoding, forced first actions, self-verification) is largely solved; squeezing frontier-grade judgment out of a small model is not.
- On-device inference systems. Wring cloud-adjacent quality from a consumer laptop — quantization that keeps a model sharp, KV-cache reuse, speculative decoding, grammar- constrained sampling, flash attention, thermal- and memory-aware scheduling. Milliseconds and megabytes are the whole game.
- Trust without a server. Safety that lives in the harness — consent, per-step undo, an append-only ledger, blocklists on resolved actions — not in hoping the model is smart. How do you make an agent you can hand your computer to when you can't audit the weights in a datacenter, because there is no datacenter?
- One brain, every device. The same logic on macOS, iPhone, Android, Windows, and Linux, proven byte-identical by machine-enforced parity tests. Portability is a correctness problem, not a packaging one.
- Honest evaluation. Measuring agentic reliability on-device, per model, per task — so "which model should I run?" is answered by data, not vibes.
We're hiring Researchers
We're looking for researchers and research-engineers who want to make small, local models genuinely capable — and who'd rather ship something a person runs today than publish a number nobody can reproduce. Areas we're hiring into:
- Inference & model efficiency — quantization, speculative decoding, KV-cache and attention systems, getting more out of every gigabyte and every watt.
- Small-model agentic capability — planning, tool-use reliability, constrained decoding, recovery behavior, and the fine-tuning / data work behind them.
- On-device evaluation — building the benchmarks that tell the truth about what a 4B can and can't do as an agent.
What we value: honest measurement (we publish every failure mode we find), shipping over posturing, and open-source by default. What we're not: we're not training a 400-billion-parameter model in a datacenter — that's precisely the thing we're routing around. If your dream is a bigger cluster, we're the wrong lab. If it's making the model in someone's pocket punch three weight classes above itself, read on.
How to apply
Email quenderin@aulenor.com with the subject “Research.” Tell us what you've built (a link beats a résumé), and which of the problems above you'd want to own. A human reads it — no portal, no tracker, no auto-reply. Remote-first; the work is judged by what it ships.