Redis Creator antirez Hand-Wrote a DeepSeek 4 Flash Metal Inference Engine in C

Salvatore Sanfilippo (known as antirez) is building again.

This time it's not Redis—it's a project called ds4: a DeepSeek 4 Flash local inference engine written in C, targeting Apple Silicon's Metal GPU acceleration. Pushed to GitHub just 13 hours ago, already at 846 stars, and it's on Hacker News.

antirez doesn't need much introduction in developer circles. Redis's original author, a C language veteran who later left Redis Labs to work on interesting smaller projects (like LoRa-related work). His coding style has two hallmarks: minimalism and speed.

What ds4 Is

From the repository structure, ds4 is remarkably lean:

• ds4.c / ds4.h: Core inference engine
• ds4_cli.c: Command-line interface
• metal/: Metal GPU acceleration kernel code
• tests/: Tests
• download_model.sh: Model download script
• AGENT.md: Instruction file for AI coding agents

Only 7 commits so far, but the quality is solid. There's already a PR fixing README typos, and antirez himself made an F32 compatibility fix for the MTP (Multi-Token Prediction) validator.

The repo includes an AGENT.md file—instructions for AI coding assistants like Claude Code and Cursor. This means antirez is already using AI agents to assist with development.

Why It Matters

First, things antirez writes are typically high quality. He's not the "write a demo, push to GitHub, and forget" type. Redis became an industry standard largely because of his obsession with code quality and performance. ds4, though newly released, already has tests, CI, and issue responsiveness—these are the signs of a seriously maintained project.

Second, Apple Silicon local inference is still early stage. While tools like llama.cpp, MLX, and Rapid MLX exist, there aren't many Metal-native inference engines specifically optimized for DeepSeek 4 Flash. If ds4 can outperform general-purpose frameworks, that's good news for Mac users.

Third, 7 commits in 13 hours—this is a project just emerging. Watching now means watching it grow from zero. If antirez keeps investing, ds4 could become the go-to solution for running DeepSeek 4 Flash on Apple Silicon.

Getting Started

# Clone the project
git clone https://github.com/antirez/ds4.git
cd ds4

# Download the model (run the script)
./download_model.sh

# Build
make

# Run
./ds4_cli

Prerequisite: you need an Apple Silicon Mac. The project depends on Metal—no Linux or Windows support.

Risks and Watch Points

It's too early to evaluate ds4. A 7-commit project is far from "production ready." Several questions to watch:

• Performance data: No public benchmarks yet. How does it compare to llama.cpp and MLX?
• Model support: Currently only supports DeepSeek 4 Flash. Will it expand?
• Maintenance cadence: antirez has a history of "wrote Redis, then moved on to other things." How long will ds4 be maintained?

I'll keep following this project. If antirez maintains his update pace, ds4 deserves a spot in the Mac local inference toolkit.

Related reading:

• Rapid MLX: Apple Silicon Inference 4x Ollama Speed
• OpenAI GPT-OSS 20B Local Deployment on MacBook

---

Primary sources:

• GitHub: antirez/ds4
• Hacker News discussion