Curate the Signal
Filter, deduplicate and decontaminate training data so every token carries information and no benchmark leaks in.
Precision over Scale
Intelligence isn't about scale; it's about precision. Pluto AI Research Lab builds compact language, code and vision-language models — 494M to 4.5B parameters — that deliver specialist capability on laptops, single GPUs and Apple Silicon, with no cloud dependency and no data leaving the device.
The Real Cost of Scale-First AI
Frontier models are remarkable — and expensive to run, hard to audit and dependent on someone else's servers. Every request is a network round-trip, a metered token and a copy of your data leaving the building. For most real work — reading a chart, completing a function, answering from a document — that trade isn't necessary. We treat model size as a cost to minimize, not a goal to maximize.
The Hidden Cost of "Bigger Is Better": a systemic inefficiency.
It forces everyday tasks to run at the speed — and the price — of a data center. For years, the model has accepted that more capability means more parameters, more GPUs and more cloud. The scale-first approach is built on waste:
No Cloud Lock-In
Closed APIs change behavior without notice and bill by the token. Open weights you run yourself can't be deprecated, repriced or rate-limited out from under you.
No Black-Box Evaluation
A single leaderboard score can hide regressions in instruction-following, verbosity and reasoning. We measure capability and behavior separately — and publish both.
No Data Exhaust
Every cloud request sends prompts, documents and code to someone else's servers. Local inference keeps them on the device that created them.
Precision Through Signal
Pluto replaces brute-force scale with signal density. We treat model quality as a data problem, not a compute arms race.
By curating every training sample, adapting strong open base models with QLoRA and distilling verified frontier behavior, we compress specialist capability into models small enough to run anywhere. Our pipeline lets us:
Distill the Behavior
Transfer reasoning and coding behavior from frontier traces into compact models with parameter-efficient fine-tuning.
Verify, Then Ship
Score models on standard benchmarks, check behavior with llm-diff, then publish GGUF and MLX builds.
Engineered for Your Hardware
When you build on Pluto, you own the whole stack — weights, data and recipes.
01
Zero Marginal Cost
Once downloaded, a Pluto model answers for free — no per-token billing, no rate limits, no usage tiers.
02
Private by Design
Prompts, documents and code stay on your device. Offline and air-gapped deployments work out of the box.
03
Reproducible by Default
Published weights, datasets and training settings mean every result can be checked, rerun and improved.
We are the bridge between frontier intelligence and the hardware you already own.
Specialist intelligence shouldn't need a data center. This isn't a smaller copy of someone else's model. It's engineered precision.