DeepSeek Models: Reasoning AI You Can Run Locally
DeepSeek made headlines with R1, a reasoning model that matches GPT-4 level thinking. While the full 671B V3 model needs a server farm, distilled versions from 7B to 70B bring strong reasoning and coding to your local hardware. DeepSeek R1 distilled models are among the best choices if you need chain-of-thought problem solving on a Mac.
DeepSeek AI4 local models
DEVELOPER
DeepSeek AI
MODELS
4
SIZE RANGE
7B–671B
RAM RANGE
10–400 GB
Key Features
Best-in-class reasoning with R1 models
Strong coding performance
Mixture-of-Experts architecture (V3)
Distilled versions run locally on modest hardware
All DeepSeek Models
| Model | Size | Quant | VRAM | Min RAM | Best For | Quality | Ollama |
|---|---|---|---|---|---|---|---|
| DeepSeek-R1 Distill Qwen 7B | 7B | Q4_K_M | 5.5 GB | 10 GB | Reasoning, Coding | 85 | |
| DeepSeek-R1 Distill Qwen 14B | 14B | Q4_K_M | 11 GB | 22 GB | Reasoning, Quality | 92 | |
| DeepSeek-R1 Distill Llama 70B | 70B | Q4_K_M | 42 GB | 48 GB | Reasoning, Quality | 97 | |
| DeepSeek-R1 671B | 671B | Q4_K_M | 380 GB | 400 GB | Reasoning, Coding | 100 |
Device Compatibility
Which DeepSeek models can run on each device class, based on minimum RAM requirements.
| Model | iPhone | Air | Pro | Studio | Mini |
|---|---|---|---|---|---|
| DeepSeek-R1 Distill Qwen 7B (7B) | Possible | Possible | Excellent | Excellent | Excellent |
| DeepSeek-R1 Distill Qwen 14B (14B) | No | Possible | Possible | Excellent | Possible |
| DeepSeek-R1 Distill Llama 70B (70B) | No | No | Possible | Possible | Possible |
| DeepSeek-R1 671B (671B) | No | No | No | Possible | No |
RAM Requirements
5.5 GB · min 10 GB
11 GB · min 22 GB
42 GB · min 48 GB
380 GB · min 400 GB
Frequently Asked Questions
Can I run DeepSeek R1 locally?
Yes. The distilled versions (7B, 14B, 32B, 70B) run locally with Ollama. The 7B distill needs 10GB RAM and is a great reasoning model for laptops.
What is the difference between DeepSeek R1 and V3?
R1 is a reasoning-focused model trained with reinforcement learning. V3 is a massive 671B MoE general model. For local use, R1 distilled models are the practical choice.
How much RAM does DeepSeek R1 need?
DeepSeek R1 7B distill needs 10GB RAM. The 14B distill needs 16GB, the 32B needs 24GB, and the 70B needs 48GB. All use Q4 quantization.
Is DeepSeek R1 good for coding?
Excellent. DeepSeek R1 distilled models show strong coding performance, especially at 14B and above. The chain-of-thought reasoning helps with complex debugging and architecture questions.