Sparse MoE: 47B total but only ~13B active per token, so it decodes fast — if it fits in memory.
chatgeneralreasoning · sizes: 8x7B (MoE) · Apache-2.0
| machine | best fit | verdict | speed |
|---|---|---|---|
| Apple M1 · 8GB | 8x7B (MoE) Q3_K_M | ❌ Won't fit | — |
| Apple M1 Pro · 16GB | 8x7B (MoE) Q3_K_M | ❌ Won't fit | — |
| Apple M2 · 16GB | 8x7B (MoE) Q3_K_M | ❌ Won't fit | — |
| Apple M3 Pro · 18GB | 8x7B (MoE) Q3_K_M | ❌ Won't fit | — |
| Apple M2 Pro · 32GB | 8x7B (MoE) Q4_K_M | 🐢 CPU-only (slow) | 2–3 tok/s |
| Apple M4 Pro · 48GB | 8x7B (MoE) Q4_K_M | ✅ Runs (tight) | 16–25 tok/s |
| Apple M3 Max · 36GB | 8x7B (MoE) Q4_K_M | 🐢 CPU-only (slow) | 3–4 tok/s |
| Apple M2 Max · 64GB | 8x7B (MoE) Q5_K_M | ✅ Runs comfortably | 18–30 tok/s |
| Apple M4 Max · 128GB | 8x7B (MoE) Q8_0 | ✅ Runs comfortably | 16–25 tok/s |
| Apple M2 Ultra · 192GB | 8x7B (MoE) Q8_0 | ✅ Runs comfortably | 17–30 tok/s |
| NVIDIA GeForce RTX 3060 · 12GB VRAM | 8x7B (MoE) Q3_K_M | 🐢 CPU-only (slow) | — |
| NVIDIA GeForce RTX 4070 · 12GB VRAM | 8x7B (MoE) Q3_K_M | 🐢 CPU-only (slow) | — |
| NVIDIA GeForce RTX 4090 · 24GB VRAM | 8x7B (MoE) Q4_K_M | ⚠️ Partial GPU offload | 15–25 tok/s |
| NVIDIA GeForce RTX 5080 · 16GB VRAM | 8x7B (MoE) Q3_K_M | 🐢 CPU-only (slow) | — |
| NVIDIA GeForce RTX 5090 · 32GB VRAM | 8x7B (MoE) Q4_K_M | ✅ Runs (tight) | 85–140 tok/s |
| AMD Radeon RX 7900 XTX · 24GB VRAM | 8x7B (MoE) Q4_K_M | ⚠️ Partial GPU offload | 13–20 tok/s |
Shopping for a machine? What hardware do I need to run Mixtral 8x7B? →
Check your exact rig: open the checker →
19 models share Mixtral 8x7B's KV-cache geometry — they size memory the same way, so a rig that fits one tends to fit its same-size kin.
…and 7 more.
Explore the whole catalog by architecture: open the graph →