Autor: Jakub Rusinowski · Ostatnia aktualizacja: 30 lipca 2026
These are the strongest local models that fit entirely in 12 GB of VRAM, ranked by capability, with the quantization level and estimated tokens/sec needed to fit.
| Qwen 2.5 Family — Qwen 2.5 14B Instruct | Q4_K_M · 9.5 GB · ~48 tok/s on Intel Arc B580 |
| Qwen 3 — Qwen 3 14B | Q4_K_M · 9.5 GB · ~48 tok/s on Intel Arc B580 |
| DeepSeek R1 — DeepSeek R1 Distill Qwen 14B | Q4_K_M · 9.2 GB · ~50 tok/s on Intel Arc B580 |
| Phi-4 Family — Phi-4 (14B) | Q4_K_M · 9.2 GB · ~50 tok/s on Intel Arc B580 |
| Qwen3-Coder — Qwen3-Coder 80B-A3B (MoE) | Q4_K_M · 7.5 GB · ~400 tok/s on Intel Arc B580 |
| Qwen 3 — Qwen 3 30B-A3B (MoE) | Q4_K_M · 8 GB · ~400 tok/s on Intel Arc B580 |
| Granite 3.0 — Granite 3.0 8B Instruct | Q4_K_M · 5.5 GB · ~83 tok/s on Intel Arc B580 |
| Qwen 3.5 (Legacy Listing — Unverified) — Qwen 3.5 14B | Q4_K_M · 9.5 GB · ~48 tok/s on Intel Arc B580 |
| Bonsai 27B — Ternary Bonsai 27B | Ternary (1.58-bit, ~1.71 bpw) · 5.9 GB · ~77 tok/s on Intel Arc B580 |
| Qwen 2.5 Family — Qwen 2.5 7B Instruct | Q4_K_M · 4.8 GB · ~95 tok/s on Intel Arc B580 |
| Qwen 3 — Qwen 3 8B | Q4_K_M · 5.5 GB · ~83 tok/s on Intel Arc B580 |
| DeepSeek R1 — DeepSeek R1 Distill Llama 8B | Q4_K_M · 5.8 GB · ~79 tok/s on Intel Arc B580 |
| Qwen3-Coder — Qwen3-Coder 8B | Q4_K_M · 5.5 GB · ~83 tok/s on Intel Arc B580 |
| Mistral Family — Mistral NeMo 12B | Q4_K_M · 8.5 GB · ~54 tok/s on Intel Arc B580 |
| Qwen 3.5 (Legacy Listing — Unverified) — Qwen 3.5 14B | Q4_K_M · 9 GB · ~51 tok/s on Intel Arc B580 |
Qwen 2.5 Family, Qwen 3, DeepSeek R1, Phi-4 Family, Qwen3-Coder all fit in 12 GB VRAM.
Intel Arc B580, NVIDIA GeForce RTX 3060 (12GB), NVIDIA GeForce RTX 5070, NVIDIA GeForce RTX 4070 Super.