{
 "id": "llama-3.1-405b",
 "name": "Llama 3.1 405B",
 "repo": "meta-llama/Llama-3.1-405B",
 "lab": {
  "name": "Meta",
  "country": "United States"
 },
 "released": "2024-07-16",
 "parameters": 405.85,
 "activeParameters": null,
 "maxContext": 131072,
 "architecture": "126 layers · GQA, 8 KV heads × 128",
 "page": "https://studiotvai.com/vram-requirements/llama-3-1-405b",
 "calculator": "https://studiotvai.com/llm-vram-calculator?model=llama-3.1-405b",
 "weightsGiB": {
  "bf16": 755.95,
  "fp8": 381.89,
  "q4_k_m": 228.63
 },
 "kvCacheGiBPerRequest": [
  {
   "context": 8192,
   "bf16": 3.938,
   "fp8": 1.969
  },
  {
   "context": 32768,
   "bf16": 15.75,
   "fp8": 7.875
  },
  {
   "context": 131072,
   "bf16": 63,
   "fp8": 31.5
  }
 ],
 "recurrentStateGiBPerRequest": 0,
 "gpusNeeded": {
  "scenario": "one request of 8192 tokens, BF16 KV cache; vLLM on data-center GPUs, llama.cpp elsewhere",
  "gpus": [
   {
    "gpu": "rtx-4090",
    "name": "GeForce RTX 4090",
    "memoryGB": 24,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 48,
     "fp8": 24,
     "q4_k_m": 16
    }
   },
   {
    "gpu": "rtx-5090",
    "name": "GeForce RTX 5090",
    "memoryGB": 32,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 32,
     "fp8": 16,
     "q4_k_m": 16
    }
   },
   {
    "gpu": "rtx-pro-6000",
    "name": "RTX PRO 6000 Blackwell",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 16,
     "fp8": 5,
     "q4_k_m": 3
    }
   },
   {
    "gpu": "mac-m5-max-128",
    "name": "Mac M5 Max 128 GB",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 16,
     "fp8": 5,
     "q4_k_m": 3
    }
   },
   {
    "gpu": "h100",
    "name": "NVIDIA H100",
    "memoryGB": 80,
    "engine": "vLLM",
    "byFormat": {
     "bf16": 16,
     "fp8": 6,
     "q4_k_m": 4
    }
   },
   {
    "gpu": "h200",
    "name": "NVIDIA H200",
    "memoryGB": 141,
    "engine": "vLLM",
    "byFormat": {
     "bf16": 7,
     "fp8": 4,
     "q4_k_m": 2
    }
   },
   {
    "gpu": "b200",
    "name": "NVIDIA B200 (HGX)",
    "memoryGB": 180,
    "engine": "vLLM",
    "byFormat": {
     "bf16": 5,
     "fp8": 3,
     "q4_k_m": 2
    }
   }
  ]
 },
 "ggufFiles": {
  "repo": "ThomasBaruzier/Meta-Llama-3.1-405B-Instruct-GGUF",
  "gib": {
   "F16": 755.98,
   "IQ1_M": 87.08,
   "IQ1_S": 79.38,
   "IQ2_M": 127.28,
   "IQ2_S": 117.02,
   "IQ2_XS": 111.15,
   "IQ2_XXS": 99.91,
   "IQ3_M": 169.26,
   "IQ3_S": 163.46,
   "IQ3_XS": 154.66,
   "IQ3_XXS": 145.14,
   "IQ4_NL": 213.39,
   "IQ4_XS": 201.7,
   "Q2_K": 139.07,
   "Q2_K_S": 128.33,
   "Q3_K_L": 198.22,
   "Q3_K_M": 181.96,
   "Q3_K_S": 163.19,
   "Q4_K_M": 226.38,
   "Q4_K_S": 214.67
  }
 },
 "runsWith": [
  "vLLM",
  "SGLang",
  "llama.cpp (Ollama, LM Studio)"
 ],
 "verified": {
  "gguf": {
   "day": "2026-10-10",
   "repo": "ThomasBaruzier/Meta-Llama-3.1-405B-Instruct-GGUF",
   "kv": "match",
   "ours": 516096,
   "theirs": 516096
  }
 },
 "updated": "2026-10-10T15:05:26.095Z"
}