Redis founder builds mini engine to run DeepSeek V4 on a 128GB MacBook Pro
The race to run massive AI models on consumer hardware just got a new contender — and it comes from the creator of Redis.
Redis founder Salvatore Sanfilippo (antirez) has released ds4.c, a minimal local inference engine designed specifically to run DeepSeek V4 Flash on Apple Silicon Macs.
To fit a mixture-of-experts model with hundreds of billions of parameters into a laptop with 128GB of RAM, ds4.c uses an asymmetric 2-bit quantization strategy — crushing the routing expert layers to extreme compression while keeping the shared expert and projection layers at higher Q8 precision. Another key innovation: it persists KV cache directly to SSD. When using a code agent to handle long-context projects, subsequent conversations can read the disk state directly, skipping the notoriously slow prefilling phase.
In real-world testing, the engine achieves about 26 tokens per second on a MacBook Pro M3 Max with 128GB of memory. By abandoning platform generality — there's no support for Nvidia GPUs at all — and optimizing for a specific hardware-software combination, Sanfilippo proves that a single hacker can still push cutting-edge AI workflows through consumer-grade hardware without being entirely beholden to massive cloud GPU clusters.