Embla 1: A reasoning model built for phones - Sevren

Sevren has released Embla 1, a reasoning model designed to run locally, privately, and offline on mobile phones. The model utilizes a custom architecture featuring sparse attention, shared K/V caches, and an engram table to achieve fast execution and low token usage per answer. Tested at 4 bits on an iPhone 18 Pro, Embla 1 averages 57.4 across several benchmarks and responds in 1.77 seconds end-to-end, though it scores lower on math benchmarks compared to some competing small models. Open article →