Antirez Runs DeepSeek V4 Flash on a MacBook

Salvatore Sanfilippo, the creator of Redis, released ds4, a native inference engine that runs DeepSeek V4 Flash locally on 128GB MacBooks using Metal acceleration. It hit 392 points on Hacker News.

The key innovation is treating the compressed KV cache as a first-class disk citizen. Instead of trying to fit the entire 1M-token context window in RAM, ds4 uses the MacBook's fast NVMe SSD as a tier of the memory hierarchy. Only MoE experts get 2-bit quantization. Shared experts and projections stay at full precision. The result: a frontier-class model running locally with a million-token context window on hardware you can buy at the Apple Store.

This matters because of who built it and how. Antirez is a systems programmer's systems programmer. ds4.c is a single C file. No framework. No abstraction layers. No dependency tree. Just a legendary engineer solving inference with the same systems thinking he brought to Redis. The approach is opinionated: treat the problem as a systems problem, not an ML problem, and the answer looks different.

The connection to the control flow story is direct. Great tools come from great engineering applied to specific constraints, not from more prompts or bigger models. Antirez looked at the problem of running a 284B-parameter model on consumer hardware and asked "what if the SSD is just another memory tier?" That's a systems insight, and it produces a tool that prompt engineering never could.

For builders with 128GB MacBooks: try this. Running a frontier model locally changes your relationship with inference. No API costs, no rate limits, no privacy concerns. The tokens-per-second won't match cloud inference, but for many workflows like code review, document analysis, and research assistance, latency matters less than availability and cost.