Yes — DeepSeek R2 (671B) runs at 12 tok/s on M4 Ultra with 192 GB RAM using Q3_K_M quantization via MLX. First token latency is 2.0s. A capable open-source LLM with 671B parameters.
LLMCheck measured DeepSeek R2 on M4 Ultra using the standard methodology: Q4_K_M quantization, 256-token input, 512-token output, 3 runs averaged on a freshly-booted system.
| Metric | Value |
|---|---|
| Tokens per second | 12 tok/s |
| Time to first token | 2.0s |
| Quantization | Q3_K_M |
| Minimum RAM | 192 GB |
| Recommended engine | MLX |
| Parameters | 671B |
| Benchmark date | 2026-05 |
Q3_K_M 671B MLX M4 Ultra
The recommended engine for DeepSeek R2 on M4 Ultra is MLX. Install with pip and pull the model:
Alternatively, you can use Ollama for a simpler setup:
To run DeepSeek R2 on M4 Ultra you need:
See how DeepSeek R2 stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard