{
 "id": "hy4-flash",
 "name": "Hy4 Flash",
 "repo": "tencent/Hy4-flash",
 "lab": {
  "name": "Tencent",
  "country": "China"
 },
 "released": "2026-11-17",
 "parameters": 220.25,
 "activeParameters": 11.3,
 "maxContext": 1048576,
 "architecture": "MoE 512 latent experts + 1 shared · MLA (512 + 64) on 42 layers: 12 sparse-attention layers (indexer keys per 4-token block) + 30 sliding-window layers (512 tokens) · 63B Engram tables in RAM · image input",
 "page": "https://studiotvai.com/vram-requirements/hy4-flash",
 "calculator": "https://studiotvai.com/llm-vram-calculator?model=hy4-flash",
 "weightsGiB": {
  "native": 266.45,
  "bf16": 408.67,
  "fp8": 205.03,
  "q4_k_m": 123.18
 },
 "kvCacheGiBPerRequest": [
  {
   "context": 8192,
   "bf16": 0.125,
   "fp8": 0.072
  },
  {
   "context": 32768,
   "bf16": 0.45,
   "fp8": 0.262
  },
  {
   "context": 131072,
   "bf16": 1.752,
   "fp8": 1.019
  },
  {
   "context": 1048576,
   "bf16": 13.903,
   "fp8": 8.084
  }
 ],
 "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": {
     "native": 16,
     "bf16": 16,
     "fp8": 7,
     "q4_k_m": 4
    }
   },
   {
    "gpu": "rtx-5090",
    "name": "GeForce RTX 5090",
    "memoryGB": 32,
    "engine": "llama.cpp",
    "byFormat": {
     "native": 7,
     "bf16": 16,
     "fp8": 5,
     "q4_k_m": 3
    }
   },
   {
    "gpu": "rtx-pro-6000",
    "name": "RTX PRO 6000 Blackwell",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "native": 3,
     "bf16": 4,
     "fp8": 2,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "mac-m5-max-128",
    "name": "Mac M5 Max 128 GB",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "native": 3,
     "bf16": 4,
     "fp8": 2,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "strix-halo",
    "name": "Ryzen AI Max+ 395 128 GB",
    "memoryGB": 96,
    "engine": "llama.cpp",
    "byFormat": {
     "native": 3,
     "bf16": 4,
     "fp8": 2,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "dgx-spark",
    "name": "DGX Spark (GB10)",
    "memoryGB": 128,
    "engine": "llama.cpp",
    "byFormat": {
     "native": 2,
     "bf16": 3,
     "fp8": 2,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "h100",
    "name": "NVIDIA H100",
    "memoryGB": 80,
    "engine": "vLLM",
    "byFormat": {
     "native": 3,
     "bf16": 5,
     "fp8": 3,
     "q4_k_m": 2
    }
   },
   {
    "gpu": "h200",
    "name": "NVIDIA H200",
    "memoryGB": 141,
    "engine": "vLLM",
    "byFormat": {
     "native": 2,
     "bf16": 3,
     "fp8": 2,
     "q4_k_m": 1
    }
   },
   {
    "gpu": "b200",
    "name": "NVIDIA B200 (HGX)",
    "memoryGB": 180,
    "engine": "vLLM",
    "byFormat": {
     "native": 2,
     "bf16": 2,
     "fp8": 1,
     "q4_k_m": 1
    }
   }
  ]
 },
 "ggufFiles": {
  "repo": "AngelSlim/Hy4-flash-GGUF",
  "gib": {
   "IQ2_AH": 125.73,
   "Q3_K": 144.7,
   "Q4_K_M": 156.79
  }
 },
 "runsWith": [
  "SGLang",
  "MLX (Mac)"
 ],
 "verified": null,
 "updated": "2026-11-22T00:05:05.824Z"
}
