Can You Run a Local LLM on an Xperia Phone?

Xperia Model Compatibility

Four current Xperia models support local LLM inference, with capability determined entirely by RAM and chipset — the new Xperia 1 VIII leads, the Xperia 1 VI remains a strong second tier, the Xperia 5 V covers the mid-range, and the Xperia 10 VI is limited to the smallest models. Sony does not pre-install on-device AI assistants on any of them (unlike Samsung Galaxy AI), so local LLM apps are the only route to private, offline AI on Xperia devices.

The Xperia 1 VIII, released June 2026 with a Snapdragon 8 Elite Gen 5 SoC and up to 16 GB of LPDDR5X RAM, is now the most capable Xperia for local inference. Its extra RAM over the Xperia 1 VI gives it comfortable headroom to run Phi-4 Q4 (~8 GB) and Qwen3-8B Q4 (~5 GB) at the same time, alongside normal app usage. The Xperia 1 VI (Snapdragon 8 Gen 3, 12 GB RAM) is still capable of quantized 7B+ models — Phi-4 Q4 and Qwen3-3B run reliably, just with less multitasking headroom than the VIII. The Xperia 5 V with Snapdragon 8 Gen 2 and 8 GB RAM is the sweet spot for 3B models — Rinna 3.6B and Gemma 2B Q4 run reliably. The Xperia 10 VI uses the mid-range Snapdragon 6 Gen 1 with only 6 GB RAM; at this tier, stick to TinyLlama 1.1B or Gemma 2B — larger models will crash or OOM during loading.

Use Xperia 1 VIII or Xperia 1 VI for 7B+ models; use Xperia 5 V for 3B models; stick to sub-2B models on Xperia 10 VI.

For app setup instructions, see our Android LLM apps for Japan guide.

3-Step Setup Guide

Installing a local LLM on an Xperia takes three steps and under 30 minutes, including model download time. The process requires no root access, no developer mode, and no special Xperia settings — it runs entirely through standard Android app and file management.

Step 1: Install MLC Chat or PocketPal AI from Google Play (Google Playストア). Both are free and available in Japan without a VPN or region workaround. MLC Chat is faster to first inference; PocketPal AI supports a broader range of GGUF model files from Hugging Face.

Step 2: Download your model over Wi-Fi. Model download sizes vary: TinyLlama 1.1B Q4 is approximately 0.7 GB, Rinna 3.6B Q4 is approximately 2 GB, Gemma 2B Q4 is approximately 1.5 GB, Phi-4 Q4 is approximately 8 GB, and Qwen3-8B Q4 is approximately 5 GB. Use a 128 GB or larger storage Xperia for Phi-4. Close all other apps before loading Phi-4 Q4 — it uses approximately 8 GB of the Xperia 1 VI's 12 GB RAM (or the Xperia 1 VIII's 16 GB) and requires maximum available memory to load without crashing. Do not download over mobile data — the files are large and your carrier plan will not thank you.

Step 3: Switch your keyboard to Japanese input. Gboard with Japanese enabled or ATOK (popular in Japan for business use) both work directly with MLC Chat and PocketPal AI — you type in Japanese, the model responds in Japanese. No extra configuration is required for Japanese language input to function.

Battery note: expect approximately 15% battery drain per hour with screen on during continuous inference on the Xperia 1 VI or Xperia 1 VIII — the newer chipset improves efficiency, but sustained on-device inference still taxes the battery at a similar rate. Enable airplane mode (機内モード) during inference sessions to reduce background radio drain and extend session time. Power-saving mode further reduces drain but may throttle the Snapdragon's AI cores and slow inference speed.

Sony Xperia AI Agent (currently in beta) connects to cloud AI services and does not run on-device. Local LLM via MLC Chat is the only way to run AI inference entirely on the Xperia without sending data to external servers — an important distinction for privacy under Japan's Act on the Protection of Personal Information (APPI / 個人情報保護法). For a full guide to Android LLM setup including hardware requirements, see running AI on tablets and Android phones.