Phone vs PC for local LLMs: why fit beats raw power

A recent XDA piece made the observation that the author’s iPhone runs local LLMs faster than their gaming PC. It sounds wrong until you actually use both setups for real work.

The desktop is faster on paper. The phone is faster in context. And context is where most LLM work actually happens.

The desktop is stretched, not slow

My desktop has a discrete GPU with finite VRAM, LM Studio open, a browser, Figma, Obsidian, and whatever else I need. If the 9B model I want to use does not fit entirely in VRAM, some layers get pushed to system RAM and CPU, and generation speed collapses. To keep the machine usable, I leave headroom, which means fewer GPU layers, which means slower tokens. The hardware is not the bottleneck — the multi-purpose workload is.

The phone is already focused

On a phone the model is the app. There is no browser battle for VRAM, no GPU layer slider to tune, no ten-second load because the app remembers the last model. You open it, type, and tokens appear. First-token latency often beats cloud chat because there is no network round-trip at all.

Modern small models are also built for this. A 0.5B–1B Q4_K_M model is not a brute-forced desktop weights file; it is a smartphone deployment target. Quality has compressed faster than the parameter count suggests.

This applies to Android too — without GPU

The XDA example uses an iPhone with Metal and unified memory. That is one path. LlamaBox takes a different one: CPU-default on Android. No GPU dependency means it runs on mid-range devices, old flagships, and budget phones that have no usable compute driver path at all. The trade-off is smaller models and modest tok/s, but the fit is the same: the device you already have, the workflow you are already in, no setup tax.

Honest limits

• Context length fills up fast on small models
• Sustained generation can thermal-throttle
• Heavy reasoning or document parsing still belongs on a desktop with RAM to spare

The real comparison

The phone does not beat the PC at everything. It beats the PC at the small, frequent, interruptible tasks that make up most LLM use: rephrase this, summarize that, draft a reply, check this claim. The tool you reach for is the one that removes friction, not the one with the best benchmark.

LlamaBox is built for that reach. Download a small GGUF, load it once, and the model stays in your pocket — no gaming PC required.

← All posts