作者： Jakub Rusinowski · 最后更新： 2026年7月30日
These are the strongest local models that fit entirely in 48 GB of VRAM, ranked by capability, with the quantization level and estimated tokens/sec needed to fit.
| Qwen 2.5 Family — Qwen 2.5 Coder 32B | Q4_K_M · 19.5 GB · ~49 tok/s on NVIDIA RTX 6000 Ada Generation |
| Llama 3.3 — Llama 3.3 70B Instruct | Q2_K_XS (Tight) · 26 GB · ~37 tok/s on NVIDIA RTX 6000 Ada Generation |
| Qwen 3 — Qwen 3 32B | Q4_K_M · 19.5 GB · ~49 tok/s on NVIDIA RTX 6000 Ada Generation |
| DeepSeek R1 — DeepSeek R1 Distill Qwen 32B | Q4_K_M · 19.5 GB · ~49 tok/s on NVIDIA RTX 6000 Ada Generation |
| Kimi K2.5 / K2.6 — Kimi K2.6 | Q4_K_M · 19 GB · ~51 tok/s on NVIDIA RTX 6000 Ada Generation |
| Qwen 2.5 Family — Qwen 2.5 14B Instruct | Q4_K_M · 9.5 GB · ~101 tok/s on NVIDIA RTX 6000 Ada Generation |
| Kimi K2.5 / K2.6 — Kimi K2.5 | Q4_K_M · 19 GB · ~51 tok/s on NVIDIA RTX 6000 Ada Generation |
| Nemotron 70B — Nemotron 70B Instruct | Q4_K_M · 39 GB · ~25 tok/s on NVIDIA RTX 6000 Ada Generation |
| Qwen 3.5 (Legacy Listing — Unverified) — Qwen 3.5 122B-A10B (MoE) | Q4_K_M · 13.5 GB · ~400 tok/s on NVIDIA RTX 6000 Ada Generation |
| Gemma 4 (Legacy Listing — Unverified) — Gemma 4 27B ⭐ | Q4_K_M · 14 GB · ~69 tok/s on NVIDIA RTX 6000 Ada Generation |
| Qwen 3 — Qwen 3 14B | Q4_K_M · 9.5 GB · ~101 tok/s on NVIDIA RTX 6000 Ada Generation |
| Qwen 3.7 — Qwen 3.7 35B-A3B | Q4_K_M · 21 GB · ~400 tok/s on NVIDIA RTX 6000 Ada Generation |
| Llama 4 — Llama 4 Maverick 17B | Q4_K_M · 24 GB · ~400 tok/s on NVIDIA RTX 6000 Ada Generation |
| Codestral — Codestral 22B | Q4_K_M · 13 GB · ~74 tok/s on NVIDIA RTX 6000 Ada Generation |
| Qwen 3.6 — Qwen 3.6 35B-A3B | Q4_K_M · 21 GB · ~400 tok/s on NVIDIA RTX 6000 Ada Generation |
Qwen 2.5 Family, Llama 3.3, Qwen 3, DeepSeek R1, Kimi K2.5 / K2.6 all fit in 48 GB VRAM.
NVIDIA RTX 6000 Ada Generation, NVIDIA L40S, Apple M5 Pro, Apple M3 Max.