The short version: certain wavelengths of red and near-infrared light pass through skin and are absorbed by structures inside the cells underneath. The most studied absorber is cytochrome c oxidase, an enzyme that sits at the end of the mitochondrial respiratory chain — the part of the cell that produces usable energy.
The leading hypothesis is that light absorption at that site briefly changes how efficiently the enzyme works, which nudges cellular energy production and sets off a cascade of secondary signals. That cascade — not the light itself — is what researchers are actually measuring when they study circulation, inflammatory markers or tissue repair.
Light is the input. Everything interesting happens several steps downstream, which is exactly why effects take weeks rather than minutes.
Two properties decide whether any of this happens in practice. Wavelength determines how deep the light reaches: around 660nm is absorbed close to the surface, while 850nm penetrates further into tissue. Dose — irradiance multiplied by time — determines whether you land in the range where anything measurable occurs.
Dose is where most consumer devices fall down, in both directions. Too little and nothing happens. Too much and the effect can plateau or reverse, a pattern researchers call the biphasic dose response. It is the single most important idea on this page, and the reason a device with four times the LEDs is not four times better.