MODEL / META

Llama 3.2 3B

Long advertised context still increases KV-cache memory; use modest context on small devices.

Primaire bron

Totale parameters3B

Geschat geheugen~2.8 GBGGUF Q4

Context128KText

LicenseLlama 3.2 Community8 official languages

01 / GESCHIKT VOOR

chat
summaries
mobile

02 / GESCHATTE FIT

Mac Studio M5 Ultra · 96 GBRuim Mac mini M6 · 24 GBRuim MacBook Pro M5 Max · 64 GBRuim Framework Desktop · Ryzen AI Max+ 395 · 128 GBRuim Radeon AI PRO R9700 · 32 GBRuim

03 / DRAAI MET

These are runtime examples, not validated installation recipes. Fit estimates assume quantized files; official repositories may hold full-precision weights. MLX, GGUF and TensorRT require matching conversions or engines.

OllamaEasyollama run llama3.2:3b

llama.cppMediumllama-server -m /path/to/verified-model.gguf --port 9931

LM StudioEasySearch “meta-llama/Llama-3.2-3B-Instruct” → Load a Q4/INT4 build → Start Server

MLX LMMediummlx_lm.server --model meta-llama/Llama-3.2-3B-Instruct

Let op

Long advertised context still increases KV-cache memory; use modest context on small devices.