What to Buy for Local Inference? antirez's Hardware Lab Breakdown

antirez — the creator of Redis — has been spending most of his time building Dwarf Star (ds4), a native Metal inference engine for DeepSeek V4 Flash. And he's been testing it on a lot of hardware.

In a recent video, he walked through what's worth buying for local AI inference right now.

1. Apple M3 Ultra (512GB) — Best, but discontinued

The absolute best machine. 512GB unified memory lets it run DeepSeek Pro at 100-200 tok/s prefill and ~11 tok/s generation. Only machine that can run frontier models at usable speeds. Cost: ~€10,000-12,000. No longer sold.

2. Apple M5 Max (128GB) — Best price/performance

DeepSeek Flash: ~400 tok/s prefill, ~37 tok/s generation. DeepSeek Pro: 3-4 tok/s. The sweet spot for local inference today. Cost: ~€6,000-7,000.

3. DGX Spark — Mixed

400 tok/s prefill but only ~15 tok/s generation due to slow memory. Costs €5,000. Has serious thermal issues — needs air conditioning. Not recommended at this price.

4. Framework Desktop (Strix Halo, AMD) — Best budget option

~€3,000-3,500 with 128GB. DeepSeek Flash: 200 tok/s prefill, 15 tok/s generation. DeepSeek Pro: only 1.6 tok/s (too slow). Runs Linux, good form factor. Decent for single-user inference + dev server.

The bottom line

For most people: the M5 Max with 128GB is the best buy right now.

If you don't want macOS: the Framework Desktop is the best alternative — slower but much cheaper and runs Linux natively.

The dream machine: If Apple releases an M5 Ultra, it would reach ~800 tok/s prefill and ~50 tok/s generation.

Underlying concern: a growing "hardware divide" where hyperscalers consume all GPU supply, leaving consumers with limited and expensive options.

This post was adapted from antirez's video "What to buy for local inference?" — transcript analyzed by Silicon.