Run DeepSeek V4 Flash on Apple Silicon
DeepSeek V4 Flash is a 284 billion parameter mixture of experts model with 13B active per token and a 1M token context, released by DeepSeek on 24 April 2026 under the MIT license. All 284B weights have to sit in memory, about 172 GB at Q4_K_M, so a Mac Studio M3 Ultra with 256 GB is the single Mac that runs it. Only 13B parameters are read for each token, so it generates much faster than a dense model of its size.
How much memory does DeepSeek V4 Flash need?
Weight size is parameter count × bits per weight; recommended memory adds 15% runtime headroom plus a KV-cache allowance for the 1,048,576-token context window.
| Quantization | Weights | Recommended RAM | Cheapest fit | Notes |
|---|---|---|---|---|
| Q4_K_M | ~172.2 GB | 216 GB | M3 Ultra (Mac Studio) | Best quality/size trade-off — the default for most deployments |
| Q8_0 | ~301.8 GB | 360 GB | Thunderbolt 5 cluster | Near-lossless; noticeably better on code and long reasoning |
| FP16 | ~568 GB | 672 GB | Thunderbolt 5 cluster | Full precision — research and eval baselines |
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.
| Chip | Max RAM | Bandwidth | Fits DeepSeek V4 Flash? | From |
|---|---|---|---|---|
| M4 · Mac mini | 32 GB | 120 GB/s | Cluster only | $99/mo |
| M4 Pro · Mac mini | 64 GB | 273 GB/s | Cluster only | $200/mo |
| M4 Max · Mac Studio | 128 GB | 546 GB/s | Cluster only | $286/mo |
| M3 Ultra · Mac Studio | 256 GB | 819 GB/s | Q4 | $571/mo |
Deploy DeepSeek V4 Flash in about 5 minutes
- 1
Build a Mac in the constructor — pick an M3 Ultra with at least 216 GB of unified memory.
- 2
Pick any Ollama-based stack (or a clean machine) and pull DeepSeek V4 Flash with one command — every deployment ships the OpenAI-compatible API.
- 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-v4-flash",
messages=[{"role": "user", "content": "Hello!"}],
)License and what it's good at
MIT License — MIT license: commercial use is allowed; keep the copyright and license notice. Made by DeepSeek.
- MIT license, with no use based restrictions
- Mixture of experts: 13B active parameters per token
- 1M token context window
Typical workloads: private coding and agent backends on one machine; long document analysis with a very large context; a local replacement for hosted deepseek apis.
Frequently asked questions
Can a Mac run DeepSeek V4 Flash?
Yes. A Mac Studio M3 Ultra with 256 GB holds the Q4_K_M build, about 172 GB of weights, with room left for context. A 128 GB Mac does not hold it at Q4.
How big is the official download?
DeepSeek's release on Hugging Face is about 160 GB. GGUF builds for llama.cpp vary with the bits per weight you choose.
Why does a model with 13B active parameters need so much memory?
Each token can be routed to different experts, so all 284B parameters have to be loaded. Active parameters set the speed; total parameters set the memory.