{
 "id": "gemma-3-12b",
 "name": "Gemma 3 12B",
 "repo": "google/gemma-3-12b-it",
 "lab": {
  "name": "Google",
  "country": "United States"
 },
 "released": "2025-03-12",
 "parameters": 12.19,
 "activeParameters": null,
 "maxContext": 131072,
 "architecture": "8 full-attention layers + 40 sliding-window layers (1024 tokens), GQA 8 KV × 256 · vision encoder",
 "page": "https://studiotvai.com/vram-requirements/gemma-3-12b",
 "calculator": "https://studiotvai.com/llm-vram-calculator?model=gemma-3-12b",
 "weightsGiB": {
  "bf16": 22.71,
  "fp8": 12.29,
  "q4_k_m": 6.82
 },
 "kvCacheGiBPerRequest": [
  {
   "context": 8192,
   "bf16": 0.813,
   "fp8": 0.406
  },
  {
   "context": 32768,
   "bf16": 2.313,
   "fp8": 1.156
  },
  {
   "context": 131072,
   "bf16": 8.313,
   "fp8": 4.156
  }
 ],
 "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": 2,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "rtx-5090",
    "name": "GeForce RTX 5090",
    "memoryGB": 32,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "rtx-pro-6000",
    "name": "RTX PRO 6000 Blackwell",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "mac-m5-max-128",
    "name": "Mac M5 Max 128 GB",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "strix-halo",
    "name": "Ryzen AI Max+ 395 128 GB",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "dgx-spark",
    "name": "DGX Spark (GB10)",
    "memoryGB": 128,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "h100",
    "name": "NVIDIA H100",
    "memoryGB": 80,
    "engine": "vLLM",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "h200",
    "name": "NVIDIA H200",
    "memoryGB": 141,
    "engine": "vLLM",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "b200",
    "name": "NVIDIA B200 (HGX)",
    "memoryGB": 180,
    "engine": "vLLM",
    "byFormat": {
     "bf16": 1,
     "fp8": 1,
     "q4_k_m": 1
    }
   }
  ]
 },
 "ggufFiles": {
  "repo": "unsloth/gemma-3-12b-it-GGUF",
  "gib": {
   "BF16": 21.92,
   "IQ4_NL": 6.41,
   "IQ4_XS": 6.1,
   "Q2_K": 4.44,
   "Q2_K_L": 4.44,
   "Q3_K_M": 5.6,
   "Q3_K_S": 5.08,
   "Q4_0": 6.43,
   "Q4_1": 7.04,
   "Q4_K_M": 6.8,
   "Q4_K_S": 6.46,
   "Q5_K_M": 7.87,
   "Q5_K_S": 7.67,
   "Q6_K": 9,
   "Q8_0": 11.65,
   "UD-IQ1_M": 3.03,
   "UD-IQ1_S": 2.85,
   "UD-IQ2_M": 4.07,
   "UD-IQ2_XXS": 3.36,
   "UD-IQ3_XXS": 4.5,
   "UD-Q2_K_XL": 4.53,
   "UD-Q3_K_XL": 5.72,
   "UD-Q4_K_XL": 6.92,
   "UD-Q5_K_XL": 7.88,
   "UD-Q6_K_XL": 9.83,
   "UD-Q8_K_XL": 13.41
  }
 },
 "runsWith": [
  "vLLM",
  "SGLang",
  "llama.cpp (Ollama, LM Studio)",
  "MLX (Mac)"
 ],
 "verified": null,
 "updated": "2026-11-06T21:02:26.190Z"
}