{
 "id": "ernie-4.5-21b-a3b",
 "name": "ERNIE 4.5 21B-A3B",
 "repo": "baidu/ERNIE-4.5-21B-A3B-Thinking",
 "lab": {
  "name": "Baidu",
  "country": "China"
 },
 "released": "2025-09-08",
 "parameters": 21.825437888,
 "activeParameters": null,
 "maxContext": 131072,
 "architecture": "28 layers · 28 full attention (4 KV × 128)",
 "page": "https://studiotvai.com/vram-requirements/ernie-4-5-21b-a3b",
 "calculator": "https://studiotvai.com/llm-vram-calculator?model=ernie-4.5-21b-a3b",
 "weightsGiB": {
  "bf16": 40.65,
  "fp8": 20.57,
  "q4_k_m": 12.33
 },
 "kvCacheGiBPerRequest": [
  {
   "context": 8192,
   "bf16": 0.438,
   "fp8": 0.219
  },
  {
   "context": 32768,
   "bf16": 1.75,
   "fp8": 0.875
  },
  {
   "context": 131072,
   "bf16": 7,
   "fp8": 3.5
  }
 ],
 "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": 2,
     "fp8": 1,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "rtx-5090",
    "name": "GeForce RTX 5090",
    "memoryGB": 32,
    "engine": "llama.cpp",
    "byFormat": {
     "bf16": 2,
     "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/ERNIE-4.5-21B-A3B-Thinking-GGUF",
  "gib": {
   "BF16": 40.66,
   "IQ4_NL": 11.62,
   "IQ4_XS": 11.01,
   "Q2_K": 7.59,
   "Q2_K_L": 7.59,
   "Q3_K_M": 9.8,
   "Q3_K_S": 8.98,
   "Q4_0": 11.64,
   "Q4_1": 12.83,
   "Q4_K_M": 12.42,
   "Q4_K_S": 11.68,
   "Q5_K_M": 14.51,
   "Q5_K_S": 14.1,
   "Q6_K": 16.74,
   "Q8_0": 21.61,
   "UD-IQ1_M": 6.75,
   "UD-IQ1_S": 6.53,
   "UD-IQ2_M": 7.48,
   "UD-IQ2_XXS": 7.23,
   "UD-IQ3_XXS": 8.89,
   "UD-Q2_K_XL": 7.66,
   "UD-Q3_K_XL": 9.85,
   "UD-Q4_K_XL": 12.48,
   "UD-Q5_K_XL": 14.6,
   "UD-Q6_K_XL": 17.94,
   "UD-Q8_K_XL": 24.81,
   "UD-TQ1_0": 6.05
  }
 },
 "runsWith": [
  "vLLM",
  "SGLang",
  "llama.cpp (Ollama, LM Studio)",
  "MLX (Mac)"
 ],
 "verified": {
  "gguf": {
   "day": "2026-11-22",
   "repo": "unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF",
   "kv": "match",
   "ours": 57344,
   "theirs": 57344
  }
 },
 "updated": "2026-11-22T00:05:05.824Z"
}
