as of today, can you run deepseek on a mac studio m1 max 32gb
- User
as of today, can you run deepseek on a mac studio m1 max 32gb
Assistant
Yes, you can run DeepSeek models on your Mac Studio M1 Max 32GB, but with some important caveats:
DeepSeek Models You Can Run
-
DeepSeek-Coder (1.3B, 6.7B, 33B)
- 1.3B & 6.7B: Run smoothly on your M1 Max
- 33B: Possible with 4-bit quantization, but will be slow (~2-5 tokens/sec)
- Best for: Code generation, completion, debugging
-
DeepSeek-LLM (7B, 67B)
- 7B: Runs well on M1 Max
- 67B: Technically possible with aggressive quantization (Q4_K_M), but performance will be poor
How to Run DeepSeek on M1 Max
Option 1: Ollama (Easiest)
ollama run deepseek-coder:6.7bOption 2: llama.cpp
- Download GGUF quantized versions from Hugging Face
- Use with llama.cpp (optimized for Apple Silicon)
Option 3: MLX (Apple-optimized)
- Some community ports available, but less common
Performance Expectations
- 6.7B models: ~15-25 tokens/sec
- 33B models (Q4): ~2-5 tokens/sec
- 67B models (Q4): <1 token/sec (not practical)
Memory Usage
- 6.7B (Q4): ~4-5GB RAM
- 33B (Q4): ~18-20GB RAM
- 67B (Q4): ~35-40GB RAM (exceeds your 32GB)
Recommendation
Stick with DeepSeek-Coder 6.7B for the best balance of performance and capability on your M1 Max. The 33B version is possible but will be frustratingly slow for interactive use.
-
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.