Open reasoning model. Distilled into Qwen/Llama backbones you can actually run locally.
reasoningcodinggeneral · sizes: 1.5B, 7B, 8B (Llama), 14B, 32B, 70B (Llama) · MIT
| machine | best fit | verdict | speed |
|---|---|---|---|
| Apple M1 · 8GB | 1.5B Q8_0 | ✅ Runs (tight) | 19–30 tok/s |
| Apple M1 Pro · 16GB | 8B (Llama) Q5_K_M | ✅ Runs (tight) | 16–25 tok/s |
| Apple M2 · 16GB | 8B (Llama) Q5_K_M | ✅ Runs (tight) | 9–14 tok/s |
| Apple M3 Pro · 18GB | 14B Q3_K_M | ✅ Runs (tight) | 9–16 tok/s |
| Apple M2 Pro · 32GB | 32B Q3_K_M | ✅ Runs (tight) | 6–10 tok/s |
| Apple M4 Pro · 48GB | 70B (Llama) Q2_K | ✅ Runs (tight) | 5–8 tok/s |
| Apple M3 Max · 36GB | 32B Q3_K_M | ✅ Runs (tight) | 10–17 tok/s |
| Apple M2 Max · 64GB | 70B (Llama) Q4_K_M | ✅ Runs (tight) | 4–7 tok/s |
| Apple M4 Max · 128GB | 70B (Llama) Q8_0 | ✅ Runs (tight) | 3–5 tok/s |
| Apple M2 Ultra · 192GB | 70B (Llama) Q8_0 | ✅ Runs comfortably | 3–5 tok/s |
| NVIDIA GeForce RTX 3060 · 12GB VRAM | 14B Q4_K_M | ✅ Runs (tight) | 14–25 tok/s |
| NVIDIA GeForce RTX 4070 · 12GB VRAM | 14B Q4_K_M | ✅ Runs (tight) | 20–35 tok/s |
| NVIDIA GeForce RTX 4090 · 24GB VRAM | 32B Q4_K_M | ✅ Runs (tight) | 19–30 tok/s |
| NVIDIA GeForce RTX 5080 · 16GB VRAM | 14B Q5_K_M | ✅ Runs (tight) | 35–55 tok/s |
| NVIDIA GeForce RTX 5090 · 32GB VRAM | 70B (Llama) Q2_K | ✅ Runs (tight) | 25–40 tok/s |
| AMD Radeon RX 7900 XTX · 24GB VRAM | 32B Q4_K_M | ✅ Runs (tight) | 17–30 tok/s |
Shopping for a machine? What hardware do I need to run DeepSeek-R1 (distill)? →
Check your exact rig: open the checker →
23 models share DeepSeek-R1 (distill)'s KV-cache geometry — they size memory the same way, so a rig that fits one tends to fit its same-size kin.
…and 11 more.
Explore the whole catalog by architecture: open the graph →