Run DeepSeek V3 on Apple Silicon

DeepSeek V3 is a 671-billion-parameter Mixture-of-Experts model with 37B active per token. MoE routing makes generation faster than a dense model of this size, but ALL expert weights must sit in memory: a Q4 build is roughly 400 GB. On Apple Silicon that means a Thunderbolt 5 cluster — two M3 Ultra 256 GB nodes pooling unified memory is the realistic entry configuration.

Parameters
671B MoE (37B active)
Min memory (Q4_K_M)
480 GB
From
TB5 cluster

How much memory does DeepSeek V3 need?

Weight size is parameter count × bits per weight; recommended memory adds 15% runtime headroom plus a KV-cache allowance for the 131,072-token context window.

QuantizationWeightsRecommended RAMCheapest fitNotes
Q4_K_M~406.8 GB480 GBThunderbolt 5 clusterBest quality/size trade-off — the default for most deployments
Q8_0~712.9 GB832 GBThunderbolt 5 clusterNear-lossless; noticeably better on code and long reasoning
FP16~1342 GB1560 GBThunderbolt 5 clusterFull precision — research and eval baselines

Realistic configs: 2 × M3 Ultra 256 GB over Thunderbolt 5 (Q4 with headroom). A single 256 GB machine cannot hold the expert weights — smaller DeepSeek distills (R1 8B/70B) exist for single-node use.

Which Mac should you pick?

Generation speed on Apple Silicon scales mostly with memory bandwidth, so the chip choice sets both whether the model fits and how fast it runs.

ChipMax RAMBandwidthFits DeepSeek V3?From
M4 · Mac mini32 GB120 GB/sCluster only$99/mo
M4 Pro · Mac mini64 GB273 GB/sCluster only$199/mo
M4 Max · Mac Studio128 GB546 GB/sCluster only$286/mo
M3 Ultra · Mac Studio256 GB819 GB/sCluster only$572/mo

Deploy DeepSeek V3 in about 5 minutes

  1. 1

    Build a Mac in the constructor — pick your cluster config with at least 480 GB of unified memory.

  2. 2

    Select the DeepSeek V3 stack from the catalog — Ollama runtime, weights pre-loaded, OpenAI-compatible API out of the box.

  3. 3

    Point your existing OpenAI SDK code at the new base_url and ship.

from openai import OpenAI

client = OpenAI(
    api_key="mcy_live_...",
    base_url="https://dep-xxxx.macyou.cloud/v1",
)
response = client.chat.completions.create(
    model="deepseek-v3",
    messages=[{"role": "user", "content": "Hello!"}],
)

License and what it's good at

DeepSeek License Model license permits commercial use with use-based restrictions; review the DeepSeek Model License for your case. Made by DeepSeek.

Typical workloads: maximum-quality private reasoning and code generation; distillation teacher for smaller models; frontier evals without sending data to third-party apis.

Frequently asked questions

Can a single Mac run DeepSeek V3?

No — Q4 weights are ~400 GB and the largest single machine (M3 Ultra) tops out at 256 GB. You need a Thunderbolt 5 cluster pooling memory across nodes, or a smaller DeepSeek distill (R1 8B / R1 70B) that fits one machine.

Why does an MoE model need all 671B parameters in RAM if only 37B are active?

The router picks different experts for every token, so any expert can be needed at any moment. Active parameters set the compute per token (speed); total parameters set the memory footprint. That is why V3 generates relatively fast once loaded, but loading it takes datacenter-class memory — or pooled unified memory on a Mac cluster.

What is the cheapest way to get DeepSeek-quality output on one Mac?

Run a DeepSeek R1 distill: R1 8B fits a 16 GB M4 (we measured 20 tokens/sec — see our benchmarks page), and R1 70B-class distills fit a 64 GB M4 Pro at Q4.

Related