CALIBRADO CON CONFIG.JSON DE HUGGINGFACE Y BENCHMARKS DE LLAMA.CPP
Calculadora Exacta de VRAM, KV-Cache y Velocidad para LLMs
Calcula al milímetro la memoria GPU requerida (Pesos + Embeddings + KV Cache GQA/MLA + Buffer CUDA), velocidad en tokens/seg y costo de alquiler Cloud GPU para 36 modelos y 23 GPUs.
⚡ Comparador Rápido Modelo × GPU
🔥 Benchmarks Más Buscados de Compatibilidad LLM × GPU
FITS (Q4_K_M)
~35 tok/s
DeepSeek-R1-Distill-Qwen-32B
on NVIDIA RTX 4090 (24GB)
VRAM: 21.23 GB (Weights 18.73G + KV 2G) →
FITS (Q4_K_M)
~35 tok/s
Qwen3-32B
on NVIDIA RTX 4090 (24GB)
VRAM: 21.22 GB (Weights 18.73G + KV 2G) →
FITS (Q4_K_M)
~103.1 tok/s
DeepSeek-R1 671B (MoE)
on 8x NVIDIA H100 80GB (640GB) (640GB)
VRAM: 378.7 GB (Weights 377.67G + KV 0.54G) →
FITS (Q4_K_M)
~27.6 tok/s
DeepSeek-R1 671B (MoE)
on Apple M3 Ultra Mac Studio (512GB) (512GB)
VRAM: 378.7 GB (Weights 377.67G + KV 0.54G) →
FITS (Q4_K_M)
~15.1 tok/s
Llama-3.3-70B-Instruct
on 2x NVIDIA RTX 3090 (48GB) (48GB)
VRAM: 43.09 GB (Weights 40.09G + KV 2.5G) →
FITS (Q4_K_M)
~32.9 tok/s
Llama-3.3-70B-Instruct
on NVIDIA A100 80GB SXM (80GB)
VRAM: 43.09 GB (Weights 40.09G + KV 2.5G) →
FITS (Q4_K_M)
~343.5 tok/s
Qwen3-30B-A3B (MoE)
on NVIDIA RTX 4090 (24GB)
VRAM: 18.53 GB (Weights 17.29G + KV 0.75G) →
FITS (Q4_K_M)
~62.3 tok/s
Qwen2.5-Coder-32B-Instruct
on NVIDIA RTX 5090 (32GB)
VRAM: 21.23 GB (Weights 18.73G + KV 2G) →
FITS (Q4_K_M)
~41.9 tok/s
Google Gemma-3-27B-IT
on NVIDIA RTX 4090 (24GB)
VRAM: 20.34 GB (Weights 15.97G + KV 3.88G) →
FITS (Q4_K_M)
~134.9 tok/s
Llama-4-Scout-17B-16E (109B MoE)
on NVIDIA A100 80GB SXM (80GB)
VRAM: 63.7 GB (Weights 61.71G + KV 1.5G) →
FITS (Q4_K_M)
~35.5 tok/s
Mistral-Small-3.1-24B-Instruct
on NVIDIA RTX 4080 Super (16GB)
VRAM: 15.28 GB (Weights 13.54G + KV 1.25G) →
FITS (Q4_K_M)
~27.7 tok/s
Qwen3-14B
on NVIDIA RTX 3060 12GB (12GB)
VRAM: 10.36 GB (Weights 8.62G + KV 1.25G) →
📊 Tabla de Referencia de Memoria VRAM (Contexto 8K)
| Modelo | Parámetros (Activos) | Arquitectura | VRAM FP16 | FP8 / Q8_0 | Q4_K_M (4-bit) | GPU Mínima (4-bit) |
|---|---|---|---|---|---|---|
| DeepSeek-R1 671B (MoE) | 671B (37B) | MLA-MoE | 1250.87 GB | 645.53 GB | 378.7 GB | Apple M3 Ultra Mac Studio (512GB) (512GB) |
| DeepSeek-V3-0324 671B (MoE) | 671B (37B) | MLA-MoE | 1250.87 GB | 645.53 GB | 378.7 GB | Apple M3 Ultra Mac Studio (512GB) (512GB) |
| DeepSeek-R1-Distill-Llama-70B | 70.55B (70.55B) | Dense-GQA | 134.41 GB | 70.82 GB | 43.09 GB | NVIDIA RTX 6000 Ada (48GB) |
| DeepSeek-R1-Distill-Qwen-32B | 32.76B (32.76B) | Dense-GQA | 63.51 GB | 34 GB | 21.23 GB | NVIDIA RTX 4090 (24GB) |
| DeepSeek-R1-Distill-Qwen-14B | 14.77B (14.77B) | Dense-GQA | 29.5 GB | 16.22 GB | 10.61 GB | NVIDIA RTX 3060 12GB (12GB) |
| DeepSeek-R1-Distill-Llama-8B | 8.03B (8.03B) | Dense-GQA | 16.45 GB | 9.23 GB | 6.22 GB | NVIDIA RTX 3060 12GB (12GB) |
| DeepSeek-R1-Distill-Qwen-7B | 7.62B (7.62B) | Dense-GQA | 15.12 GB | 8.28 GB | 5.43 GB | NVIDIA RTX 3060 12GB (12GB) |
| Qwen3-235B-A22B (MoE) | 235.1B (22.2B) | MoE-GQA | 439.87 GB | 227.79 GB | 134.4 GB | NVIDIA H200 141GB SXM (141GB) |
| Qwen3-30B-A3B (MoE) | 30.53B (3.34B) | MoE-GQA | 58.1 GB | 30.57 GB | 18.53 GB | NVIDIA RTX 4090 (24GB) |
| Qwen3-32B | 32.76B (32.76B) | Dense-GQA | 63.51 GB | 34 GB | 21.22 GB | NVIDIA RTX 4090 (24GB) |
| Qwen3-14B | 14.77B (14.77B) | Dense-GQA | 29.25 GB | 15.97 GB | 10.36 GB | NVIDIA RTX 3060 12GB (12GB) |
| Qwen3-8B | 8.19B (8.19B) | Dense-GQA | 16.87 GB | 9.52 GB | 6.47 GB | NVIDIA RTX 3060 12GB (12GB) |
| Qwen3-4B | 4.02B (4.02B) | Dense-GQA | 8.68 GB | 5.07 GB | 3.61 GB | NVIDIA RTX 3060 12GB (12GB) |
| Qwen2.5-72B-Instruct | 72.71B (72.71B) | Dense-GQA | 138.43 GB | 72.9 GB | 44.38 GB | NVIDIA RTX 6000 Ada (48GB) |
| Qwen2.5-Coder-32B-Instruct | 32.76B (32.76B) | Dense-GQA | 63.51 GB | 34 GB | 21.23 GB | NVIDIA RTX 4090 (24GB) |
| Qwen2.5-14B-Instruct | 14.77B (14.77B) | Dense-GQA | 29.5 GB | 16.22 GB | 10.61 GB | NVIDIA RTX 3060 12GB (12GB) |
| Qwen2.5-7B-Instruct | 7.62B (7.62B) | Dense-GQA | 15.12 GB | 8.28 GB | 5.43 GB | NVIDIA RTX 3060 12GB (12GB) |
| QwQ-32B Reasoning | 32.76B (32.76B) | Dense-GQA | 63.51 GB | 34 GB | 21.23 GB | NVIDIA RTX 4090 (24GB) |
| Llama-4-Maverick-17B-128E (400B MoE) | 400B (17.2B) | MoE-GQA | 747.05 GB | 386.22 GB | 227.32 GB | 4x NVIDIA A100 80GB (320GB) (320GB) |
| Llama-4-Scout-17B-16E (109B MoE) | 109B (17.2B) | MoE-GQA | 205.02 GB | 106.74 GB | 63.7 GB | Apple M4 Pro Mac Mini (64GB) (64GB) |
| Llama-3.3-70B-Instruct | 70.55B (70.55B) | Dense-GQA | 134.41 GB | 70.82 GB | 43.09 GB | NVIDIA RTX 6000 Ada (48GB) |
| Llama-3.1-405B-Instruct | 405B (405B) | Dense-GQA | 758.82 GB | 393.55 GB | 233.02 GB | 4x NVIDIA A100 80GB (320GB) (320GB) |
| Llama-3.1-70B-Instruct | 70.55B (70.55B) | Dense-GQA | 134.41 GB | 70.82 GB | 43.09 GB | NVIDIA RTX 6000 Ada (48GB) |
| Llama-3.1-8B-Instruct | 8.03B (8.03B) | Dense-GQA | 16.45 GB | 9.23 GB | 6.22 GB | NVIDIA RTX 3060 12GB (12GB) |
| Llama-3.2-3B-Instruct | 3.21B (3.21B) | Dense-GQA | 7.34 GB | 4.47 GB | 3.33 GB | NVIDIA RTX 3060 12GB (12GB) |
| Mistral-Large-2-123B | 122.6B (122.6B) | Dense-GQA | 231.62 GB | 121.1 GB | 72.84 GB | NVIDIA A100 80GB SXM (80GB) |
| Mistral-Small-3.1-24B-Instruct | 23.6B (23.6B) | Dense-GQA | 45.7 GB | 24.45 GB | 15.28 GB | NVIDIA RTX 5080 (16GB) |
| Mixtral-8x22B-Instruct (MoE) | 140.6B (39.1B) | MoE-GQA | 264.13 GB | 137.29 GB | 81.38 GB | Apple M4 Max (128GB Unified) (128GB) |
| Mixtral-8x7B-Instruct (MoE) | 46.7B (12.9B) | MoE-GQA | 88.47 GB | 46.35 GB | 27.8 GB | NVIDIA RTX 5090 (32GB) |
| Codestral-22B | 22.2B (22.2B) | Dense-GQA | 43.59 GB | 23.58 GB | 14.81 GB | NVIDIA RTX 5080 (16GB) |
| Google Gemma-3-27B-IT | 27.4B (27.4B) | Dense-GQA | 55.4 GB | 30.76 GB | 20.34 GB | NVIDIA RTX 4090 (24GB) |
| Google Gemma-3-12B-IT | 12.2B (12.2B) | Dense-GQA | 26.21 GB | 15.26 GB | 10.75 GB | NVIDIA RTX 3060 12GB (12GB) |
| Google Gemma-3-4B-IT | 4.3B (4.3B) | Dense-GQA | 9.56 GB | 5.72 GB | 4.24 GB | NVIDIA RTX 3060 12GB (12GB) |
| Microsoft Phi-4 14B | 14.66B (14.66B) | Dense-GQA | 29.36 GB | 16.16 GB | 10.5 GB | NVIDIA RTX 3060 12GB (12GB) |
| Cohere Command-A 111B | 111B (111B) | Dense-GQA | 209.26 GB | 109.3 GB | 66.19 GB | NVIDIA A100 80GB SXM (80GB) |
| GLM-4-32B-0414 | 32B (32B) | Dense-GQA | 62 GB | 33.19 GB | 20.77 GB | NVIDIA RTX 4090 (24GB) |