Research note ·
Gearshift: Can One Model Think and Another Finish?
What I learned trying to hand reasoning between models without replaying it
Investigation 01 / AI inference
Ongoing researchGearshift.
Can inference move between models?
An investigation into heterogeneous inference, including the possibility of handing off a KV cache from one model to another.
Gearshift / research questionConceptual
01Model AInference stateKV cache
State handoff?
?
An open question02Model BContinue inference?Different model
The question
What would it take to continue somewhere else?
Gearshift asks whether an inference process could begin with one model and continue with another. The investigation includes KV-cache handoff: whether the state accumulated during inference can be made useful across model boundaries.
The work is ongoing. We’re examining the conditions that a handoff would need to satisfy, how to evaluate it, and what would make it useful. This overview makes no claim of a successful transfer.
Lines of inquiry
01 / Compatibility
What can carry over?
What constraints would a different model place on the state it receives?
02 / Evaluation
How would we know?
How should continuity and output quality be tested against a clear baseline?
03 / Usefulness
When would it help?
What costs and tradeoffs would determine whether a handoff is worth making?