Meta released Llama 4 edge checkpoints tuned for phones and laptops, targeting offline inference with 4-bit quantization and sub-2B active-parameter variants for battery-constrained devices.
Running locally
The weights ship under Meta’s open license with reference runtimes for Android, iOS, and desktop via popular local runners. Meta claims parity within a few points of cloud APIs on summarization and lightweight coding tasks.
On-device privacy is the headline: sensitive dictation, field notes, and CRM snippets can stay local while still benefiting from a capable chat model.
Sensitive dictation and field notes can stay on-device — no token bill, no upload queue.
Developers get updated fine-tuning notebooks aimed at domain adapters small enough to fit beside the base edge checkpoint in unified memory.
