chatreasoningcodingagents · sizes: 20B-A3.6B (MoE), 120B-A5.1B (MoE)
| machine | best fit | verdict | speed |
|---|---|---|---|
| Apple M1 · 8GB | 120B-A5.1B (MoE) MXFP4 | ❌ Won't fit | — |
| Apple M1 Pro · 16GB | 20B-A3.6B (MoE) MXFP4 | 🐢 CPU-only (slow) | 3–5 tok/s |
| Apple M2 · 16GB | 20B-A3.6B (MoE) MXFP4 | 🐢 CPU-only (slow) | 2–3 tok/s |
| Apple M3 Pro · 18GB | 20B-A3.6B (MoE) MXFP4 | 🐢 CPU-only (slow) | 2–4 tok/s |
| Apple M2 Pro · 32GB | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 25–40 tok/s |
| Apple M4 Pro · 48GB | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 30–50 tok/s |
| Apple M3 Max · 36GB | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 40–70 tok/s |
| Apple M2 Max · 64GB | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 40–70 tok/s |
| Apple M4 Max · 128GB | 120B-A5.1B (MoE) MXFP4 | ✅ Runs comfortably | 35–60 tok/s |
| Apple M2 Ultra · 192GB | 120B-A5.1B (MoE) MXFP4 | ✅ Runs comfortably | 50–85 tok/s |
| NVIDIA GeForce RTX 3060 · 12GB VRAM | 20B-A3.6B (MoE) MXFP4 | ⚠️ Partial GPU offload | 10–17 tok/s |
| NVIDIA GeForce RTX 4070 · 12GB VRAM | 20B-A3.6B (MoE) MXFP4 | ⚠️ Partial GPU offload | 14–25 tok/s |
| NVIDIA GeForce RTX 4090 · 24GB VRAM | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 95–160 tok/s |
| NVIDIA GeForce RTX 5080 · 16GB VRAM | 20B-A3.6B (MoE) MXFP4 | ✅ Runs (tight) | 85–140 tok/s |
| NVIDIA GeForce RTX 5090 · 32GB VRAM | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 170–280 tok/s |
| AMD Radeon RX 7900 XTX · 24GB VRAM | 20B-A3.6B (MoE) MXFP4 | ✅ Runs comfortably | 85–140 tok/s |
Shopping for a machine? What hardware do I need to run gpt-oss? →
Check your exact rig: open the checker →
One model shares gpt-oss's KV-cache geometry — they size memory the same way, so a rig that fits one tends to fit its same-size kin.
Explore the whole catalog by architecture: open the graph →