Compression-aware training
Train with the deployment target in mind
Training recipes anticipate quantization, pruning, and memory limits during training rather than treating compression as a final packaging step applied after the fact.
Research direction
Edge models
Training and compression for on-device inference.
Edge models is Rkive's research direction for training and compressing powerful multimodal systems for on-device inference. The goal is not merely smaller models — it is multimodal capability that remains useful once hardware limits are treated as first-class constraints.
Capability under device budgets
Latency, memory, thermals, and power are not afterthoughts — they are part of the optimization target from the start. This means training recipes that anticipate compression, evaluation that measures real device behavior, and distillation that preserves the cross-modal capability that makes models useful.
Multimodal distillation
Preserve capability in compact models
Compression only matters if useful cross-modal behavior survives. Distillation is aimed at keeping compact models genuinely capable rather than merely small.
Device-budget evaluation
Latency, memory, thermals, and power
Evaluation is tied to hardware budgets so that performance metrics reflect real deployment conditions, not theoretical FLOPs on unconstrained hardware.
Edge-model work applies the same disciplined search mindset to deployment reality. The same MAIC-RL protocol and structured experimentation used in AIR can be applied when the constraint shifts from artifact size to runtime budgets on real devices.