Run DeepSeek R1 Distill 8B on Apple Silicon
DeepSeek R1 Distill 8B brings o1-style chain-of-thought reasoning to a model that runs on the cheapest Apple Silicon config: we measured 20.0 tokens/sec on a base M4 with 16 GB (Q4_K_M, median of 3 runs). Budget extra tokens per answer — reasoning traces are long — but for math, logic, and stepwise analysis it punches far above its size.
How much memory does R1 8B 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.
| Quantization | Weights | Recommended RAM | Cheapest fit | Notes |
|---|---|---|---|---|
| Q4_K_M | ~4.9 GB | 16 GB | M4 (Mac mini) | Best quality/size trade-off — the default for most deployments |
| Q8_0 | ~8.5 GB | 16 GB | M4 (Mac mini) | Near-lossless; noticeably better on code and long reasoning |
| FP16 | ~16 GB | 24 GB | M4 (Mac mini) | 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 R1 8B? | From |
|---|---|---|---|---|
| M4 · Mac mini | 32 GB | 120 GB/s | Q4 + Q8 + FP16 | $99/mo |
| M4 Pro · Mac mini | 64 GB | 273 GB/s | Q4 + Q8 + FP16 | $199/mo |
| M4 Max · Mac Studio | 128 GB | 546 GB/s | Q4 + Q8 + FP16 | $286/mo |
| M3 Ultra · Mac Studio | 256 GB | 819 GB/s | Q4 + Q8 + FP16 | $572/mo |
Deploy R1 8B in about 5 minutes
- 1
Build a Mac in the constructor — pick an M4 with at least 16 GB of unified memory.
- 2
Pick any Ollama-based stack (or a clean machine) and pull DeepSeek R1 Distill 8B 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-r1-8b",
messages=[{"role": "user", "content": "Hello!"}],
)License and what it's good at
MIT (distill weights); base model licenses apply — The R1 distill weights are MIT-licensed; the Llama base carries Meta's community license terms. Made by DeepSeek.
- Real reasoning traces on a $99/mo machine — measured 20.0 tok/s
- Dramatically better at math/logic than same-size chat models
- Upgrade path to the 70B distill without changing your API code
Typical workloads: math and logic-heavy assistants; code review with visible reasoning; evaluating whether reasoning models help your workload before paying for 70b.
Frequently asked questions
Why is R1 8B slower per answer than Llama 8B at similar tok/s?
Generation speed is nearly identical (20.0 vs 21.2 tok/s measured), but R1 emits a chain of thought before the answer — often 2–5× more tokens per response. You pay in latency, not in speed.
Is there a bigger single-machine option?
Yes — the R1 Distill 70B (Llama-3.3 base) at Q4 fits a 64 GB M4 Pro. See our post “DeepSeek V3 70B doesn't exist — run the R1 Distill 70B instead” for that setup.