Msingi Labs Says Its Offline Translation Model Beats the Cloud on Latency

Msingi Labs published benchmarks this week claiming its on-device translation model returns a full sentence in 210 milliseconds on a mid-tier Android phone — roughly a third of the time it takes to reach a cloud endpoint and back from the same handset on a typical 3G connection.

The comparison is deliberately unflattering to the cloud. Msingi ran its tests on real networks in Kampala, Gulu, and Mbale rather than in a lab, and reported the 95th percentile rather than the median. Under those conditions the cloud round trip stretched past 2.4 seconds; the on-device model, unaffected by the network, did not move.

"Everyone quotes median latency, and median latency is a lie in this market," said Msingi co-founder Ronald Kizza. "The question that matters is what happens on your worst call of the day, because that is the one the user remembers."

Quality is the trade-off, and Msingi does not hide it: the model trails the leading cloud translators by about four BLEU points on formal text. The lab's argument is that for market traders, health workers, and customer support agents, a good answer now beats a better answer in three seconds — a framing that, if it holds, points toward a lot more inference moving to the edge.