按内存选择 Mac 本地 AI 模型（8GB–128GB）

The short answer: choose a Mac for local AI by unified memory first and chip generation second. Keep at least 25–30% of memory free for macOS, context, and the applications you actually use. A model that loads is not automatically a model that runs comfortably. This guide maps Mac memory tiers to realistic model sizes without treating a benchmark score as a buying recommendation. Quick recommendations 8GB: compact 2–4B models for short prompts, classification, and extraction. 16GB: general 7–9B models; 20B can run close to the limit. 24GB: the practical mainstream tier for 9–20B models. 36–48GB: the balanced range for 27–35B, coding, and document work. 64GB: 70B at 4-bit becomes practical. 96–128GB: large 70–120B models, long context, and parallel workloads. Use the AI Feed Local AI monitor to compare current models by memory, license, and task. Mac memory guide Mac memory Comfortable model class Best fit Main constraint 8GB 2–4B Q4 Short chat, extraction Very little context headroom 16GB 7–9B Q4/Q6 Writing, documents, light coding 20B runs close to the limit 24GB 14–20B; selected 27B Q4 Coding, RAG, daily assistant 27B needs moderate context 36GB 27–35B Q4 Serious coding, agents Long context still consumes memory 48GB 35B Q6; reduced 70B Large repositories, multiple agents Full 70B remains tight 64GB 70B Q4 Complex analysis and coding Smaller MoE models may be faster 96–128GB 70B high precision; 120B class Long context, serving workloads High hardware cost These are practical headroom targets, not hard limits. Actual use changes with quantization, runtime, and context length. Why memory matters more than the chip badge Apple Silicon uses unified memory: the CPU and GPU share one pool. That helps local inference because the accelerator can access most of the machine's memory without copying weights into separate VRAM. The trade-off is competition. Model weights, Safari, an IDE, Docker, the display system, and the KV cache all use the same pool. When weights consume nearly everything, macOS swaps, generation slows, and a longer prompt can terminate the process. Memory needed = weights + KV cache + compute buffers + runtime + macOS + apps Weights are the large predictable component. Q4 uses roughly half a byte per parameter, plus overhead. KV cache grows with the conversation and context window. Compute buffers depend on architecture and runtime. The system still needs room for your browser, editor, and background services. That is why a 17GB model file may run on a 24GB Mac but feel much better on a 36GB machine. What each memory tier is good for 8GB: small models only Look at 2–4B variants of Qwen, Phi, Gemma, and similar families below roughly 3–4GB. They are useful for classification, field extraction, short rewrites, and simple completion. They are not a realistic choice for a large codebase or long documents. 16GB: the entry point Modern 7–9B Q4/Q5 models can edit text, answer over local documents, and help with small functions. OpenAI says gpt-oss-20b can run within 16GB, but that leaves little room for context and applications. Possible is not the same as comfortable. 24GB: the practical mainstream choice For most buyers, this is the first genuinely useful local AI configuration. It can keep a strong 9–14B model resident alongside a browser and editor, or run gpt-oss-20b with reasonable headroom. Practical choices include: a modern 9B model for an everyday assistant; gpt-oss-20b for reasoning and agentic tasks; a 27B Q4 model as an upper-edge experiment with moderate context; a 9–20B model paired with a capable coding harness. 36GB and 48GB: the developer sweet spot Models in the 27–35B class leave room for context, an IDE, and a browser. At 48GB, higher-quality quantization or multiple processes become practical. More GPU cores make a model faster. More memory changes which class of model can run at all. For many developers, that makes 36–48GB more valuable than a faster chip paired with less memory. 64GB: practical 70B inference At 64GB, 70B models in Q4 become realistic. Their weights commonly occupy around 42–48GB, leaving limited but workable headroom. More parameters do not automatically mean a better experience. A modern 35B MoE model can be faster and more reliable with tools than an older dense 70B model. 96GB and 128GB: workstation territory These capacities make sense when you intentionally need very large models, several simultaneous clients, or long context. OpenAI estimates gpt-oss-120b at roughly 80GB, so 96GB is the practical starting point and 128GB leaves useful application headroom. For occasional demanding jobs, a cloud model can still be faster and cheaper. The AI Feed API price calculator helps compare the order of cost. Ollama, MLX, or llama.cpp? Runtime Best when Strength Trade-off Ollama You want the simplest setup Model library and local API Less tuning control MLX / MLX-LM A good Apple Silicon conversion exists Native unified-memory design New conversions may take time llama.cpp You need GGUF and detailed tuning Portability and Metal support More manual setup Start with Ollama or LM Studio. Try MLX when a reliable conversion exists. Use llama.cpp when you want direct control over context, cache, offload, and server settings. Coding, images, and video are different workloads Coding Instruction following, repository access, and tool use matter as much as model size. A good 9–20B model with a coding harness is usually more useful than a larger model with no context headroom. Images Diffusion models use memory differently from LLMs. Compact generators run at 16–24GB, while high resolution, multiple ControlNets, and batches need more. Video Local video remains the heaviest workload. A 24–36GB Mac can run compact models for short clips with long waits, but it is not equivalent to a cloud video service. A five-step buying rule Define the workload. Chat, coding, retrieval, images, and video need different hardware. Choose the model class. Start with 9–20B for daily work; not everyone needs 35B. Find the quantized file size. Check Ollama, MLX Community, or a GGUF repository. Add 25–30% headroom. Add more for long context, Docker, and a heavy IDE. Compare chips last. With equal memory, Max is faster than Pro, but it cannot run a model that does not fit. Buying recommendation 16GB for learning and small local models. 24GB as the practical minimum for regular use. 36–48GB for developers, agents, and 27–35B models. 64GB when you specifically need 70B or concurrent models. 96–128GB for a workstation or home server. Sources Apple MLX: Unified Memory llama.cpp: macOS Metal build Ollama model library OpenAI: gpt-oss memory requirements Hugging Face: gpt-oss-20b Microsoft: Phi-4 Mini Instruct Updated September 25, 2026. Sizes refer to specific quantizations and may change as runtimes evolve. Compare current models with the Local AI model selector. The original article and future updates remain on AI Feed.