Meta released a smaller Llama checkpoint aimed at on-device and edge boxes, with quantized weights and a short commercial FAQ.
Meta’s latest Llama drop is for people who count megabytes and milliseconds. The company published a compact checkpoint aimed at edge boxes and high-end phones, shipped quantized weights beside it, and added a shorter commercial FAQ so lawyers spend less time squinting at forum threads.
Where this model fits
This is not the model you pick for deep research agents. Meta’s own framing leans on latency and memory. If your product needs offline classification, short rewrites, or on-device helpers with a predictable footprint, this lane is the pitch. If you need long-horizon reasoning, stay on the large endpoints and stop pretending a tiny checkpoint will bluff its way through.
Quantized weights matter because edge deployments die on download size and RAM. Meta including them in the drop reduces the amateur hour of “we will quantize it later and hope evals still look fine.” You should still re-run your own tasks. Vendor charts measure vendor tasks.
License and ops notes
Read the FAQ that accompanies the weights. Commercial use questions are why small teams get stuck even when the model is free to download. Screenshot the FAQ with a date. Policy pages move.
On device choice, measure thermal throttling and battery, not only tokens per second on a bench PSU. An edge model that melts a phone in six minutes is a paperweight. Meta widened the menu. Your product constraints still pick the dish.
Evaluation you should run locally
Build a twenty-prompt suite that matches your product: classification, short rewrite, refusal on unsafe asks, and one multi-step instruction. Run it on the quantized file you will actually ship, not on a denser checkpoint you admired in a leaderboard screenshot. Track tokens per second after five minutes of load when the device is warm. Edge models fail in heat and in memory pressure long before they fail in a cool lab notebook. Meta handed you weights. Your suite decides if they belong in the binary.