{"id":19132,"date":"2026-03-06T12:19:42","date_gmt":"2026-03-06T12:19:42","guid":{"rendered":"https:\/\/ideainthebox.com\/?p=19132"},"modified":"2026-03-06T12:22:32","modified_gmt":"2026-03-06T12:22:32","slug":"utilizzare-un-vecchio-cellulare-come-server-ai","status":"publish","type":"post","link":"https:\/\/ideainthebox.com\/index.php\/2026\/03\/06\/utilizzare-un-vecchio-cellulare-come-server-ai\/","title":{"rendered":"Utilizzare un vecchio cellulare come SERVER AI"},"content":{"rendered":"

Ecco una guida completa per trasformare il tuo Samsung Galaxy Note 9<\/strong> in un vero server AI locale con Qwen.<\/p>\n

\n

\u26a0\ufe0f Prima cosa: realt\u00e0 sull’hardware del Note 9<\/h3>\n

Il Note 9 ha Snapdragon 845 (o Exynos 9810 in Europa), con 6 GB o 8 GB di RAM. Questo significa che puoi girare modelli piccoli e quantizzati<\/strong> (0.6B, 1.7B, max 4B) con prestazioni limitate ma funzionali. Non aspettarti GPT-4, ma per un server AI personale offline funziona.<\/p>\n

\n

\ud83d\udee0\ufe0f Metodo 1 \u2014 Ollama + Termux (pi\u00f9 semplice)<\/h3>\n

1. Installa Termux<\/h4>\n

La versione di Termux sul Google Play Store \u00e8 obsoleta \u2014 scaricala da F-Droid<\/strong> o direttamente da GitHub (github.com\/termux\/termux-app\/releases).<\/p>\n

2. Aggiorna Termux e installa Ollama<\/h4>\n
pkg update && pkg upgrade -y\r\npkg install ollama\r\n<\/code><\/pre>\n
3. Avvia il server Ollama in background<\/h4>\n
# In una sessione Termux\r\nollama serve &\r\n<\/code><\/pre>\n
4. Scarica Qwen3 (modelli consigliati per 6\/8 GB RAM)<\/h4>\n
Puoi scegliere la dimensione pi\u00f9 adatta: ollama run qwen3:0.6b<\/code> oppure ollama run qwen3:1.7b<\/code><\/p>\n
ollama pull qwen3:0.6b    # ~400 MB \u2014 veloce\r\nollama pull qwen3:1.7b    # ~1.1 GB \u2014 migliore qualit\u00e0\r\n<\/code><\/pre>\n
Con 8 GB di RAM puoi osare anche il qwen3:4b<\/code> (~2.5 GB).<\/p>\n
5. Ollama \u00e8 gi\u00e0 un server OpenAI-compatibile<\/h4>\n
Ollama espone automaticamente http:\/\/localhost:11434<\/code> con endpoint compatibili con OpenAI. Puoi testarlo con:<\/p>\n
curl http:\/\/localhost:11434\/v1\/chat\/completions \r\n  -H \"Content-Type: application\/json\" \r\n  -d '{\"model\":\"qwen3:1.7b\",\"messages\":[{\"role\":\"user\",\"content\":\"Ciao!\"}]}'\r\n<\/code><\/pre>\n
\n
\ud83d\udee0\ufe0f Metodo 2 \u2014 llama.cpp diretto (pi\u00f9 controllo)<\/h3>\n
Se vuoi pi\u00f9 controllo o usare Qwen3.5:<\/p>\n
pkg update && pkg upgrade -y\r\npkg install git cmake clang\r\n# Clona llama.cpp\r\ngit clone https:\/\/github.com\/ggml-org\/llama.cpp\r\ncd llama.cpp\r\ncmake -B build\r\ncmake --build build --config Release -j4\r\n# Scarica il modello GGUF da HuggingFace\r\npkg install wget\r\nwget https:\/\/huggingface.co\/Qwen\/Qwen3-1.7B-GGUF\/resolve\/main\/qwen3-1.7b-q4_k_m.gguf\r\n<\/code><\/pre>\n
Avvia come server HTTP (API OpenAI-compatibile):<\/h4>\n
.\/build\/bin\/llama-server \r\n  -m qwen3-1.7b-q4_k_m.gguf \r\n  --port 8080 \r\n  --ctx-size 4096 \r\n  --host 0.0.0.0\r\n<\/code><\/pre>\n
Avviando il server con --host 0.0.0.0<\/code>, potrai accedere alle sue capacit\u00e0 AI non solo dal telefono stesso ma anche da altri dispositivi sulla stessa rete, come altri app o terminali.<\/p>\n
\n
\ud83c\udf10 Renderlo accessibile dalla rete locale<\/h4>\n
Una volta avviato il server, qualsiasi device sulla stessa Wi-Fi pu\u00f2 usarlo come backend AI:<\/p>\n
http:\/\/IP_DEL_NOTE9:8080\/v1\/chat\/completions\r\n<\/code><\/pre>\n
Puoi trovare l’IP del telefono con:<\/p>\n
ip addr show wlan0\r\n<\/code><\/pre>\n
\n
\ud83d\udca1 Trucchi importanti per Android<\/h3>\n
    \n
  • Disabilita l’ottimizzazione batteria<\/strong> per Termux nelle impostazioni Android, altrimenti il processo viene killato<\/li>\n
  • Disabilita le restrizioni “phantom process” nelle opzioni sviluppatore del telefono<\/li>\n
  • Usa una microSD veloce<\/strong> per salvare i modelli (il Note 9 supporta fino a 1 TB)<\/li>\n
  • Tieni il telefono in carica durante l’uso intensivo<\/li>\n
  • Con 6 GB di RAM, chiudi tutte le altre app prima di avviare il server<\/li>\n<\/ul>\n
    \n
    \ud83d\udcca Aspettative realistiche<\/h3>\n\n\n\n\n\n\n\n
    Modello<\/th>\nRAM usata<\/th>\nToken\/sec (CPU)<\/th>\nQualit\u00e0<\/th>\n<\/tr>\n<\/thead>\n
    Qwen3-0.6B<\/td>\n~1 GB<\/td>\n8-15 t\/s<\/td>\nbase<\/td>\n<\/tr>\n
    Qwen3-1.7B<\/td>\n~2 GB<\/td>\n4-8 t\/s<\/td>\nbuona<\/td>\n<\/tr>\n
    Qwen3-4B<\/td>\n~4 GB<\/td>\n2-4 t\/s<\/td>\nottima<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
    Per Qwen3.5 (il pi\u00f9 recente), il modello da 2B \u00e8 cos\u00ec compresso da poter girare su smartphone, mentre le varianti pi\u00f9 grandi richiedono hardware pi\u00f9 potente.<\/p>\n
    #!\/data\/data\/com.termux\/files\/usr\/bin\/bash
    \n# \u2554\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2557
    \n# \u2551 NOTE 9 AI SERVER \u2014 INSTALLER v1.0 \u2551
    \n# \u2551 by Claude \u2022 Qwen3 su Termux + llama.cpp \u2551
    \n# \u255a\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u255d<\/p>\n
    # \u2500\u2500\u2500 Colori \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nRED=’33[0;31m’
    \nGREEN=’33[0;32m’
    \nYELLOW=’33[1;33m’
    \nCYAN=’33[0;36m’
    \nBOLD=’33[1m’
    \nRESET=’33[0m’<\/p>\n
    # \u2500\u2500\u2500 Variabili configurabili \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nSERVER_PORT=8080
    \nMODEL_DIR=”$HOME\/ai_models”
    \nLOG_FILE=”$HOME\/ai_server.log”
    \nSERVICE_SCRIPT=”$HOME\/start_ai_server.sh”<\/p>\n
    # \u2500\u2500\u2500 Funzioni di stampa \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nprint_banner() {
    \necho -e “${CYAN}”
    \necho ” \u2554\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2557”
    \necho ” \u2551 \ud83d\udcf1 NOTE 9 \u2014 AI SERVER INSTALLER \u2551”
    \necho ” \u2551 Qwen3 \u2022 llama.cpp \u2022 API \u2551”
    \necho ” \u255a\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u255d”
    \necho -e “${RESET}”
    \n}<\/p>\n
    step() { echo -e “n${CYAN}${BOLD}[STEP $1]${RESET} $2”; }
    \nok() { echo -e “${GREEN} \u2714 $1${RESET}”; }
    \nwarn() { echo -e “${YELLOW} \u26a0 $1${RESET}”; }
    \nerr() { echo -e “${RED} \u2717 $1${RESET}”; }
    \ninfo() { echo -e ” \u2192 $1″; }<\/p>\n
    # \u2500\u2500\u2500 Selezione modello \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nchoose_model() {
    \necho -e “n${BOLD}Scegli il modello Qwen3 da installare:${RESET}”
    \necho “”
    \necho ” [1] Qwen3-0.6B (~400 MB) \u2014 Leggero, veloce, RAM < 2 GB”
    \necho ” [2] Qwen3-1.7B (~1.1 GB) \u2014 Bilanciato \u2605 Consigliato per Note 9″
    \necho ” [3] Qwen3-4B (~2.5 GB) \u2014 Ottima qualit\u00e0, serve 8 GB RAM”
    \necho ” [4] Qwen3-8B (~5.0 GB) \u2014 Solo con RAM \u2265 8 GB + storage libero”
    \necho “”
    \nread -p ” Scelta [1-4] (default: 2): ” MODEL_CHOICE
    \nMODEL_CHOICE=${MODEL_CHOICE:-2}<\/p>\n
    case $MODEL_CHOICE in
    \n1)
    \nMODEL_SIZE=”0.6b”
    \nMODEL_FILE=”qwen3-0.6b-q4_k_m.gguf”
    \nMODEL_URL=”https:\/\/huggingface.co\/Qwen\/Qwen3-0.6B-GGUF\/resolve\/main\/qwen3-0.6b-q4_k_m.gguf”
    \n;;
    \n2)
    \nMODEL_SIZE=”1.7b”
    \nMODEL_FILE=”qwen3-1.7b-q4_k_m.gguf”
    \nMODEL_URL=”https:\/\/huggingface.co\/Qwen\/Qwen3-1.7B-GGUF\/resolve\/main\/qwen3-1.7b-q4_k_m.gguf”
    \n;;
    \n3)
    \nMODEL_SIZE=”4b”
    \nMODEL_FILE=”qwen3-4b-q4_k_m.gguf”
    \nMODEL_URL=”https:\/\/huggingface.co\/Qwen\/Qwen3-4B-GGUF\/resolve\/main\/qwen3-4b-q4_k_m.gguf”
    \n;;
    \n4)
    \nMODEL_SIZE=”8b”
    \nMODEL_FILE=”qwen3-8b-q4_k_m.gguf”
    \nMODEL_URL=”https:\/\/huggingface.co\/Qwen\/Qwen3-8B-GGUF\/resolve\/main\/qwen3-8b-q4_k_m.gguf”
    \n;;
    \n*)
    \nwarn “Scelta non valida. Uso modello 1.7B di default.”
    \nMODEL_SIZE=”1.7b”
    \nMODEL_FILE=”qwen3-1.7b-q4_k_m.gguf”
    \nMODEL_URL=”https:\/\/huggingface.co\/Qwen\/Qwen3-1.7B-GGUF\/resolve\/main\/qwen3-1.7b-q4_k_m.gguf”
    \n;;
    \nesac<\/p>\n
    ok “Modello selezionato: Qwen3-${MODEL_SIZE}”
    \n}<\/p>\n
    # \u2500\u2500\u2500 Controllo RAM disponibile \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \ncheck_ram() {
    \nTOTAL_RAM_KB=$(grep MemTotal \/proc\/meminfo | awk ‘{print $2}’)
    \nTOTAL_RAM_GB=$(echo “scale=1; $TOTAL_RAM_KB \/ 1024 \/ 1024” | bc)
    \nFREE_RAM_KB=$(grep MemAvailable \/proc\/meminfo | awk ‘{print $2}’)
    \nFREE_RAM_GB=$(echo “scale=1; $FREE_RAM_KB \/ 1024 \/ 1024” | bc)<\/p>\n
    info “RAM totale: ${TOTAL_RAM_GB} GB | RAM libera: ${FREE_RAM_GB} GB”<\/p>\n
    if [ “$TOTAL_RAM_KB” -lt 4000000 ]; then
    \nwarn “Meno di 4 GB RAM! Il modello 0.6B \u00e8 l’unica opzione sicura.”
    \nfi
    \n}<\/p>\n
    # \u2500\u2500\u2500 Controllo storage \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \ncheck_storage() {
    \nFREE_STORAGE=$(df -h “$HOME” | awk ‘NR==2{print $4}’)
    \ninfo “Storage libero in HOME: $FREE_STORAGE”
    \n}<\/p>\n
    # \u2500\u2500\u2500 Step 1: Aggiornamento pacchetti \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 1 “Aggiornamento repository Termux…”
    \npkg update -y 2>&1 | tail -3
    \npkg upgrade -y 2>&1 | tail -3
    \nok “Repository aggiornato”<\/p>\n
    # \u2500\u2500\u2500 Step 2: Dipendenze base \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 2 “Installazione dipendenze (git, cmake, clang, wget, bc)…”
    \npkg install -y git cmake clang wget bc make 2>&1 | tail -5
    \nok “Dipendenze installate”<\/p>\n
    # \u2500\u2500\u2500 Step 3: Info sistema \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 3 “Analisi hardware del dispositivo…”
    \ncheck_ram
    \ncheck_storage
    \nCPU_CORES=$(nproc)
    \ninfo “Core CPU disponibili: $CPU_CORES”<\/p>\n
    # \u2500\u2500\u2500 Step 4: Scelta modello \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 4 “Selezione modello AI…”
    \nchoose_model<\/p>\n
    # \u2500\u2500\u2500 Step 5: Compilazione llama.cpp \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 5 “Download e compilazione llama.cpp…”<\/p>\n
    if [ -d “$HOME\/llama.cpp” ]; then
    \nwarn “Cartella llama.cpp gi\u00e0 esistente. Aggiornamento…”
    \ncd “$HOME\/llama.cpp” && git pull 2>&1 | tail -3
    \nelse
    \ninfo “Clonazione repository llama.cpp…”
    \ngit clone https:\/\/github.com\/ggml-org\/llama.cpp “$HOME\/llama.cpp” 2>&1 | tail -5
    \nfi<\/p>\n
    cd “$HOME\/llama.cpp”<\/p>\n
    info “Compilazione in corso (pu\u00f2 richiedere 5-10 minuti)…”
    \ncmake -B build -DLLAMA_CURL=OFF 2>&1 | tail -3
    \ncmake –build build –config Release -j”$CPU_CORES” 2>&1 | tail -5<\/p>\n
    if [ -f “$HOME\/llama.cpp\/build\/bin\/llama-server” ]; then
    \nok “llama.cpp compilato con successo!”
    \nelse
    \nerr “Compilazione fallita. Controlla il log.”
    \nexit 1
    \nfi<\/p>\n
    # \u2500\u2500\u2500 Step 6: Download modello \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 6 “Download modello Qwen3-${MODEL_SIZE}…”<\/p>\n
    mkdir -p “$MODEL_DIR”
    \nMODEL_PATH=”$MODEL_DIR\/$MODEL_FILE”<\/p>\n
    if [ -f “$MODEL_PATH” ]; then
    \nok “Modello gi\u00e0 presente: $MODEL_PATH”
    \nelse
    \ninfo “Download da HuggingFace… (potrebbe volerci del tempo)”
    \nwget -q –show-progress -O “$MODEL_PATH” “$MODEL_URL”<\/p>\n
    if [ $? -eq 0 ]; then
    \nok “Modello scaricato: $MODEL_PATH”
    \nelse
    \nerr “Download fallito! Verifica connessione internet.”
    \nrm -f “$MODEL_PATH”
    \nexit 1
    \nfi
    \nfi<\/p>\n
    # \u2500\u2500\u2500 Step 7: Generazione script di avvio \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 7 “Creazione script di avvio server…”<\/p>\n
    cat > “$SERVICE_SCRIPT” << STARTSCRIPT
    \n#!\/data\/data\/com.termux\/files\/usr\/bin\/bash
    \n# \u2500\u2500\u2500 AI SERVER AVVIO RAPIDO \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500<\/p>\n
    MODEL_PATH=”$MODEL_PATH”
    \nSERVER_BIN=”$HOME\/llama.cpp\/build\/bin\/llama-server”
    \nPORT=$SERVER_PORT
    \nLOG=”$LOG_FILE”<\/p>\n
    echo “\ud83e\udd16 Avvio AI Server su porta $PORT…”
    \necho ” Modello: $(basename $MODEL_PATH)”
    \necho ” Endpoint: http:\/\/$(ip route get 1 | awk ‘{print $7; exit}’):$PORT”
    \necho ” API: http:\/\/localhost:$PORT\/v1\/chat\/completions”
    \necho “”
    \necho ” Premi CTRL+C per fermare il server”
    \necho “”<\/p>\n
    $SERVER_BIN
    \n–model “$MODEL_PATH”
    \n–port “$PORT”
    \n–host 0.0.0.0
    \n–ctx-size 4096
    \n–n-predict 512
    \n–threads $(nproc)
    \n–log-file “$LOG”
    \n-ngl 0<\/p>\n
    STARTSCRIPT<\/p>\n
    chmod +x “$SERVICE_SCRIPT”
    \nok “Script di avvio creato: $SERVICE_SCRIPT”<\/p>\n
    # \u2500\u2500\u2500 Step 8: Script di test \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 8 “Creazione script di test API…”<\/p>\n
    cat > “$HOME\/test_ai.sh” << ‘TESTSCRIPT’
    \n#!\/data\/data\/com.termux\/files\/usr\/bin\/bash
    \nPORT=8080
    \necho “\ud83e\uddea Test chiamata API al server locale…”
    \ncurl -s http:\/\/localhost:$PORT\/v1\/chat\/completions
    \n-H “Content-Type: application\/json”
    \n-d ‘{
    \n“model”: “qwen3”,
    \n“messages”: [{“role”:”user”,”content”:”Rispondi in italiano: chi sei e cosa puoi fare?”}],
    \n“max_tokens”: 200
    \n}’ | python3 -c ”
    \nimport sys, json
    \ntry:
    \ndata = json.load(sys.stdin)
    \nmsg = data[‘choices’][0][‘message’][‘content’]
    \nprint(‘n\u2705 Risposta AI:n’)
    \nprint(msg)
    \nexcept Exception as e:
    \nprint(‘\u274c Errore:’, e)
    \nprint(‘Raw:’, sys.stdin.read())
    \n”
    \nTESTSCRIPT<\/p>\n
    chmod +x “$HOME\/test_ai.sh”
    \nok “Script di test creato: ~\/test_ai.sh”<\/p>\n
    # \u2500\u2500\u2500 Step 9: Alias utili \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \nstep 9 “Configurazione alias rapidi in ~\/.bashrc…”<\/p>\n
    grep -q “# AI Server aliases” “$HOME\/.bashrc” 2>\/dev\/null || cat >> “$HOME\/.bashrc” << ‘ALIASES’<\/p>\n
    # AI Server aliases
    \nalias ai-start=’bash ~\/start_ai_server.sh’
    \nalias ai-test=’bash ~\/test_ai.sh’
    \nalias ai-log=’tail -f ~\/ai_server.log’
    \nalias ai-ip=’ip route get 1 | awk “{print $7; exit}”‘
    \nALIASES<\/p>\n
    ok “Alias aggiunti: ai-start, ai-test, ai-log, ai-ip”<\/p>\n
    # \u2500\u2500\u2500 Riepilogo finale \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
    \necho “”
    \necho -e “${GREEN}${BOLD}”
    \necho ” \u2554\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2557”
    \necho ” \u2551 \u2705 INSTALLAZIONE COMPLETATA! \u2551”
    \necho ” \u255a\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u255d”
    \necho -e “${RESET}”
    \necho “”
    \necho -e ” ${BOLD}COMANDI DISPONIBILI:${RESET}”
    \necho “”
    \necho -e ” ${CYAN}ai-start${RESET} \u2192 Avvia il server AI”
    \necho -e ” ${CYAN}ai-test${RESET} \u2192 Testa che il server risponda”
    \necho -e ” ${CYAN}ai-log${RESET} \u2192 Mostra i log in tempo reale”
    \necho -e ” ${CYAN}ai-ip${RESET} \u2192 Mostra l’IP del telefono”
    \necho “”
    \necho -e ” ${BOLD}ENDPOINT API (OpenAI-compatibile):${RESET}”
    \necho -e ” ${YELLOW}http:\/\/localhost:$SERVER_PORT\/v1\/chat\/completions${RESET}”
    \necho -e ” ${YELLOW}http:\/\/$(ip route get 1 2>\/dev\/null | awk ‘{print $7; exit}’ || echo ‘IP_TELEFONO’):$SERVER_PORT\/v1\/chat\/completions${RESET}”
    \necho “”
    \necho -e ” ${BOLD}AVVIO RAPIDO:${RESET}”
    \necho -e ” ${GREEN}source ~\/.bashrc && ai-start${RESET}”
    \necho “”
    \necho -e ” ${BOLD}\u26a0 IMPORTANTE:${RESET}”
    \necho -e ” Vai in Impostazioni Android \u2192 App \u2192 Termux”
    \necho -e ” \u2192 Batteria \u2192 disabilita ‘Ottimizzazione batteria’”
    \necho “”<\/p>\n
    Ecco i due file creati:<\/p>\n
    note9_ai_server_install.sh<\/code><\/strong> \u2014 Lo script bash completo che fa tutto automaticamente:<\/p>\n
      \n
    • Aggiorna i pacchetti Termux<\/li>\n
    • Installa tutte le dipendenze (git, cmake, clang, ecc.)<\/li>\n
    • Controlla la RAM disponibile<\/li>\n
    • Ti fa scegliere il modello Qwen3 (0.6B \/ 1.7B \/ 4B \/ 8B)<\/li>\n
    • Compila llama.cpp direttamente sul telefono<\/li>\n
    • Scarica il modello da HuggingFace<\/li>\n
    • Crea start_ai_server.sh<\/code> e test_ai.sh<\/code><\/li>\n
    • Aggiunge gli alias ai-start<\/code>, ai-test<\/code>, ai-log<\/code>, ai-ip<\/code><\/li>\n<\/ul>\n
       <\/p>\n
      Come usarlo sul Note 9:<\/strong><\/p>\n
        \n
      1. Copia lo script .sh<\/code> sul telefono via USB o cloud<\/li>\n
      2. Apri Termux<\/strong> (da F-Droid, non Play Store!)<\/li>\n
      3. Esegui:<\/li>\n<\/ol>\n
        bash<\/p>\n
           bash \/sdcard\/Download\/note9_ai_server_install.sh<\/code><\/pre>\n
          \n
        1. Segui le istruzioni a schermo (~10-15 min per la compilazione)<\/li>\n
        2. Poi basta digitare ai-start<\/code> per avviare il server ogni volta<\/li>\n<\/ol>\n
          Guida interattiva da aprire nel browser del Note 9, con tutti i comandi da copiare con un tap.<\/p>\n
          \ud83e\udd16<\/p>\n
          Note 9 AI Server<\/h1>\n
          Qwen3 \u2022 llama.cpp \u2022 OpenAI API<\/p>\n<\/header>\n
          Installa Termux da F-Droid per iniziare<\/p>\n
          \u2460 Scegli il modello \u26a1 Qwen3-0.6B ~400 MB \u00b7 8-15 token\/sec \u00b7 RAM min 2 GB Veloce \u2b50 Qwen3-1.7B ~1.1 GB \u00b7 4-8 token\/sec \u00b7 RAM min 3 GB Consigliato \ud83d\udd25 Qwen3-4B ~2.5 GB \u00b7 2-4 token\/sec \u00b7 RAM min 6 GB 6 GB RAM \ud83d\udc8e Qwen3-8B ~5.0 GB \u00b7 1-2 token\/sec \u00b7 RAM 8 GB Solo 8 GB<\/p>\n
          \u2461 Segui i passaggi<\/p>\n
          1 Installa Termux \u25bc \u26a0 NON usare la versione del Play Store (obsoleta). Scarica da F-Droid. \ud83c\udf10 Link download<\/p>\n
          https:\/\/f-droid.org\/packages\/com.termux\/<\/pre>\n