Coding with DeepSeek 4 on a 128GB MacBook Pro
The article details how to run DeepSeek 4 Flash, a large language model, locally on a 128GB MacBook Pro using tools like Ollama or MLX. It covers setup, model quantization, and performance benchmarks, showing that the high-end MacBook can handle the model with reasonable speed for coding tasks. The guide demonstrates local AI development without cloud dependency.
Background
- DeepSeek is a Chinese AI company whose large language models (LLMs) rival OpenAI's GPT-4 and Meta's Llama, often at much lower cost. DeepSeek-V4 (also called "DeepSeek-V4 Flash") is their latest generation release.
- Most powerful LLMs require massive cloud GPU clusters to run. The author is demonstrating that DeepSeek-V4 can run *locally* (on your own computer, no internet needed) on a high-end consumer Mac — specifically a MacBook Pro with 128GB of unified memory (RAM shared between CPU and GPU via Apple's M-series chips).
- "Running locally" matters for privacy (no data sent to a third-party server), offline reliability, and avoiding per-token API fees. It also means the model has to be heavily compressed or quantized to fit within the machine's memory.
- 128GB is an unusually large amount of RAM for a laptop (standard is 8–36GB). Even so, fitting a state-of-the-art LLM requires specialized compression techniques (quantization) that trade a small amount of output quality for dramatically reduced memory footprint.