Tested On-Device AI Models

💡 Important Hardware & Context Notice

File size represents the serialized weights on disk, not total runtime RAM usage. Loading a 1.06 GB model with a 2,048-token context buffer creates an active KV-cache requiring ~1.4 GB to 1.6 GB of resident memory. PocketAI's automatic safety watchdog monitors system memory and prevents out-of-memory crashes by capping context before thermal throttling occurs.