{
  "site": "https://neurarch.com",
  "description": "Structural comparisons of neural network architectures: layer-by-layer diff, parameter and cost deltas, and GPU fit, derived from the graphs by the same functions the app runs.",
  "count": 30,
  "plan": "https://www.neurarch.com/api/v1/plan",
  "comparisons": [
    {
      "key": "gpt2-vs-qwen3-8b",
      "a": {
        "key": "gpt2",
        "name": "GPT-2",
        "layers": 10,
        "params": 84331345
      },
      "b": {
        "key": "qwen3-8b",
        "name": "Qwen3-8B",
        "layers": 219,
        "params": 8190878080
      },
      "question": "What seven years of scaling actually changed inside the decoder.",
      "headline": "Qwen3-8B has 8.1B more parameters than GPT-2: 212 layers added, 3 removed, 6 changed.",
      "paramDelta": 8106546735,
      "layerDelta": 209,
      "added": 212,
      "removed": 3,
      "changed": 6,
      "gpuFlips": [
        {
          "gpu": "T4",
          "a": true,
          "b": false
        },
        {
          "gpu": "A100_40",
          "a": true,
          "b": false
        },
        {
          "gpu": "H100_80",
          "a": true,
          "b": false
        }
      ],
      "page": "https://neurarch.com/diff/gpt2-vs-qwen3-8b.html",
      "markdown": "https://neurarch.com/diff/gpt2-vs-qwen3-8b.md"
    },
    {
      "key": "llama3-block-vs-mixtral-block",
      "a": {
        "key": "llama3-block",
        "name": "LLaMA-3 Block",
        "layers": 8,
        "params": 702554112
      },
      "b": {
        "key": "mixtral-block",
        "name": "Mixtral MoE Block",
        "layers": 7,
        "params": 1451270144
      },
      "question": "A dense feed-forward against a mixture of experts, at the block level.",
      "headline": "Mixtral MoE Block has 749M more parameters than LLaMA-3 Block: 1 layer added, 2 removed, 2 changed.",
      "paramDelta": 748716032,
      "layerDelta": -1,
      "added": 1,
      "removed": 2,
      "changed": 2,
      "gpuFlips": [
        {
          "gpu": "T4",
          "a": true,
          "b": false
        }
      ],
      "page": "https://neurarch.com/diff/llama3-block-vs-mixtral-block.html",
      "markdown": "https://neurarch.com/diff/llama3-block-vs-mixtral-block.md"
    },
    {
      "key": "mixtral-block-vs-deepseek-v3",
      "a": {
        "key": "mixtral-block",
        "name": "Mixtral MoE Block",
        "layers": 7,
        "params": 1451270144
      },
      "b": {
        "key": "deepseek-v3",
        "name": "DeepSeek-V3",
        "layers": 370,
        "params": 666329115904
      },
      "question": "Two mixture-of-experts designs that route very differently.",
      "headline": "DeepSeek-V3 has 665B more parameters than Mixtral MoE Block: 365 layers added, 2 removed, 4 changed.",
      "paramDelta": 664877845760,
      "layerDelta": 363,
      "added": 365,
      "removed": 2,
      "changed": 4,
      "gpuFlips": [
        {
          "gpu": "A100_40",
          "a": true,
          "b": false
        },
        {
          "gpu": "H100_80",
          "a": true,
          "b": false
        }
      ],
      "page": "https://neurarch.com/diff/mixtral-block-vs-deepseek-v3.html",
      "markdown": "https://neurarch.com/diff/mixtral-block-vs-deepseek-v3.md"
    },
    {
      "key": "llama3-block-vs-deepseek-v3",
      "a": {
        "key": "llama3-block",
        "name": "LLaMA-3 Block",
        "layers": 8,
        "params": 702554112
      },
      "b": {
        "key": "deepseek-v3",
        "name": "DeepSeek-V3",
        "layers": 370,
        "params": 666329115904
      },
      "question": "What a frontier open model adds to the block everyone started from.",
      "headline": "DeepSeek-V3 has 666B more parameters than LLaMA-3 Block: 364 layers added, 2 removed, 5 changed.",
      "paramDelta": 665626561792,
      "layerDelta": 362,
      "added": 364,
      "removed": 2,
      "changed": 5,
      "gpuFlips": [
        {
          "gpu": "T4",
          "a": true,
          "b": false
        },
        {
          "gpu": "A100_40",
          "a": true,
          "b": false
        },
        {
          "gpu": "H100_80",
          "a": true,
          "b": false
        }
      ],
      "page": "https://neurarch.com/diff/llama3-block-vs-deepseek-v3.html",
      "markdown": "https://neurarch.com/diff/llama3-block-vs-deepseek-v3.md"
    },
    {
      "key": "transformer-block-vs-llama3-block",
      "a": {
        "key": "transformer-block",
        "name": "Transformer Block",
        "layers": 6,
        "params": 7087872
      },
      "b": {
        "key": "llama3-block",
        "name": "LLaMA-3 Block",
        "layers": 8,
        "params": 702554112
      },
      "question": "What changed in the transformer block between 2017 and now.",
      "headline": "LLaMA-3 Block has 695M more parameters than Transformer Block: 5 layers added, 3 removed, 2 changed.",
      "paramDelta": 695466240,
      "layerDelta": 2,
      "added": 5,
      "removed": 3,
      "changed": 2,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/transformer-block-vs-llama3-block.html",
      "markdown": "https://neurarch.com/diff/transformer-block-vs-llama3-block.md"
    },
    {
      "key": "transformer-block-vs-mamba-block",
      "a": {
        "key": "transformer-block",
        "name": "Transformer Block",
        "layers": 6,
        "params": 7087872
      },
      "b": {
        "key": "mamba-block",
        "name": "Mamba SSM Block",
        "layers": 16,
        "params": 168271976
      },
      "question": "Attention against a state-space layer for the same job.",
      "headline": "Mamba SSM Block has 161M more parameters than Transformer Block: 12 layers added, 2 removed, 4 changed.",
      "paramDelta": 161184104,
      "layerDelta": 10,
      "added": 12,
      "removed": 2,
      "changed": 4,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/transformer-block-vs-mamba-block.html",
      "markdown": "https://neurarch.com/diff/transformer-block-vs-mamba-block.md"
    },
    {
      "key": "mamba-block-vs-jamba",
      "a": {
        "key": "mamba-block",
        "name": "Mamba SSM Block",
        "layers": 16,
        "params": 168271976
      },
      "b": {
        "key": "jamba",
        "name": "Jamba",
        "layers": 51,
        "params": 13030961152
      },
      "question": "Pure state space against a hybrid that keeps some attention.",
      "headline": "Jamba has 13B more parameters than Mamba SSM Block: 42 layers added, 7 removed, 8 changed.",
      "paramDelta": 12862689176,
      "layerDelta": 35,
      "added": 42,
      "removed": 7,
      "changed": 8,
      "gpuFlips": [
        {
          "gpu": "T4",
          "a": true,
          "b": false
        },
        {
          "gpu": "A100_40",
          "a": true,
          "b": false
        },
        {
          "gpu": "H100_80",
          "a": true,
          "b": false
        }
      ],
      "page": "https://neurarch.com/diff/mamba-block-vs-jamba.html",
      "markdown": "https://neurarch.com/diff/mamba-block-vs-jamba.md"
    },
    {
      "key": "gpt2-vs-bert-base",
      "a": {
        "key": "gpt2",
        "name": "GPT-2",
        "layers": 10,
        "params": 84331345
      },
      "b": {
        "key": "bert-base",
        "name": "BERT Base",
        "layers": 9,
        "params": 31120896
      },
      "question": "Decoder-only against encoder-only, same era, same size class.",
      "headline": "BERT Base has 53M fewer parameters than GPT-2: 2 layers added, 3 removed, 5 changed.",
      "paramDelta": -53210449,
      "layerDelta": -1,
      "added": 2,
      "removed": 3,
      "changed": 5,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/gpt2-vs-bert-base.html",
      "markdown": "https://neurarch.com/diff/gpt2-vs-bert-base.md"
    },
    {
      "key": "bert-base-vs-t5-small",
      "a": {
        "key": "bert-base",
        "name": "BERT Base",
        "layers": 9,
        "params": 31120896
      },
      "b": {
        "key": "t5-small",
        "name": "T5 Small",
        "layers": 20,
        "params": 56734592
      },
      "question": "Encoder-only against encoder-decoder.",
      "headline": "T5 Small has 26M more parameters than BERT Base: 14 layers added, 2 removed, 8 changed.",
      "paramDelta": 25613696,
      "layerDelta": 11,
      "added": 14,
      "removed": 2,
      "changed": 8,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/bert-base-vs-t5-small.html",
      "markdown": "https://neurarch.com/diff/bert-base-vs-t5-small.md"
    },
    {
      "key": "qwen3-8b-vs-phi3-mini",
      "a": {
        "key": "qwen3-8b",
        "name": "Qwen3-8B",
        "layers": 219,
        "params": 8190878080
      },
      "b": {
        "key": "phi3-mini",
        "name": "Phi-3 Mini Block",
        "layers": 10,
        "params": 310288704
      },
      "question": "What gets cut to make a small model small.",
      "headline": "Phi-3 Mini Block has 7.9B fewer parameters than Qwen3-8B: 1 layer added, 210 removed, 8 changed.",
      "paramDelta": -7880589376,
      "layerDelta": -209,
      "added": 1,
      "removed": 210,
      "changed": 8,
      "gpuFlips": [
        {
          "gpu": "T4",
          "a": false,
          "b": true
        },
        {
          "gpu": "A100_40",
          "a": false,
          "b": true
        },
        {
          "gpu": "H100_80",
          "a": false,
          "b": true
        }
      ],
      "page": "https://neurarch.com/diff/qwen3-8b-vs-phi3-mini.html",
      "markdown": "https://neurarch.com/diff/qwen3-8b-vs-phi3-mini.md"
    },
    {
      "key": "llama3-block-vs-phi3-mini",
      "a": {
        "key": "llama3-block",
        "name": "LLaMA-3 Block",
        "layers": 8,
        "params": 702554112
      },
      "b": {
        "key": "phi3-mini",
        "name": "Phi-3 Mini Block",
        "layers": 10,
        "params": 310288704
      },
      "question": "A block from a large model against a whole small one.",
      "headline": "Phi-3 Mini Block has 392M fewer parameters than LLaMA-3 Block: 2 layers added, 6 changed.",
      "paramDelta": -392265408,
      "layerDelta": 2,
      "added": 2,
      "removed": 0,
      "changed": 6,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/llama3-block-vs-phi3-mini.html",
      "markdown": "https://neurarch.com/diff/llama3-block-vs-phi3-mini.md"
    },
    {
      "key": "qwen3-8b-vs-llama-4-scout",
      "a": {
        "key": "qwen3-8b",
        "name": "Qwen3-8B",
        "layers": 219,
        "params": 8190878080
      },
      "b": {
        "key": "llama-4-scout",
        "name": "Llama-4 Scout",
        "layers": 431,
        "params": 102547680704
      },
      "question": "Two 2026 open models with different attention budgets.",
      "headline": "Llama-4 Scout has 94B more parameters than Qwen3-8B: 249 layers added, 36 removed, 112 changed.",
      "paramDelta": 94356802624,
      "layerDelta": 212,
      "added": 249,
      "removed": 36,
      "changed": 112,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/qwen3-8b-vs-llama-4-scout.html",
      "markdown": "https://neurarch.com/diff/qwen3-8b-vs-llama-4-scout.md"
    },
    {
      "key": "resnet-block-vs-simple-cnn",
      "a": {
        "key": "resnet-block",
        "name": "ResNet Block",
        "layers": 7,
        "params": 73984
      },
      "b": {
        "key": "simple-cnn",
        "name": "Simple CNN",
        "layers": 7,
        "params": 804554
      },
      "question": "What the residual connection costs and what it buys.",
      "headline": "Simple CNN has 731K more parameters than ResNet Block: 4 layers added, 4 removed, 2 changed.",
      "paramDelta": 730570,
      "layerDelta": 0,
      "added": 4,
      "removed": 4,
      "changed": 2,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/resnet-block-vs-simple-cnn.html",
      "markdown": "https://neurarch.com/diff/resnet-block-vs-simple-cnn.md"
    },
    {
      "key": "resnet-block-vs-vit-b16",
      "a": {
        "key": "resnet-block",
        "name": "ResNet Block",
        "layers": 7,
        "params": 73984
      },
      "b": {
        "key": "vit-b16",
        "name": "ViT-B/16",
        "layers": 11,
        "params": 8449000
      },
      "question": "Convolution against attention for images.",
      "headline": "ViT-B/16 has 8.4M more parameters than ResNet Block: 8 layers added, 4 removed, 3 changed.",
      "paramDelta": 8375016,
      "layerDelta": 4,
      "added": 8,
      "removed": 4,
      "changed": 3,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/resnet-block-vs-vit-b16.html",
      "markdown": "https://neurarch.com/diff/resnet-block-vs-vit-b16.md"
    },
    {
      "key": "vit-b16-vs-swin-tiny",
      "a": {
        "key": "vit-b16",
        "name": "ViT-B/16",
        "layers": 11,
        "params": 8449000
      },
      "b": {
        "key": "swin-tiny",
        "name": "Swin-Tiny",
        "layers": 83,
        "params": 28260808
      },
      "question": "A flat vision transformer against a hierarchical one.",
      "headline": "Swin-Tiny has 20M more parameters than ViT-B/16: 74 layers added, 2 removed, 4 changed.",
      "paramDelta": 19811808,
      "layerDelta": 72,
      "added": 74,
      "removed": 2,
      "changed": 4,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/vit-b16-vs-swin-tiny.html",
      "markdown": "https://neurarch.com/diff/vit-b16-vs-swin-tiny.md"
    },
    {
      "key": "bert-base-vs-vit-b16",
      "a": {
        "key": "bert-base",
        "name": "BERT Base",
        "layers": 9,
        "params": 31120896
      },
      "b": {
        "key": "vit-b16",
        "name": "ViT-B/16",
        "layers": 11,
        "params": 8449000
      },
      "question": "The same transformer applied to text and to images.",
      "headline": "ViT-B/16 has 23M fewer parameters than BERT Base: 4 layers added, 2 removed, 5 changed.",
      "paramDelta": -22671896,
      "layerDelta": 2,
      "added": 4,
      "removed": 2,
      "changed": 5,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/bert-base-vs-vit-b16.html",
      "markdown": "https://neurarch.com/diff/bert-base-vs-vit-b16.md"
    },
    {
      "key": "unet-vs-diffusion-unet",
      "a": {
        "key": "unet",
        "name": "U-Net",
        "layers": 22,
        "params": 720705
      },
      "b": {
        "key": "diffusion-unet",
        "name": "Diffusion UNet",
        "layers": 17,
        "params": 6687044
      },
      "question": "A segmentation U-Net against the one a diffusion model uses.",
      "headline": "Diffusion UNet has 6.0M more parameters than U-Net: 7 layers added, 12 removed, 11 changed.",
      "paramDelta": 5966339,
      "layerDelta": -5,
      "added": 7,
      "removed": 12,
      "changed": 11,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/unet-vs-diffusion-unet.html",
      "markdown": "https://neurarch.com/diff/unet-vs-diffusion-unet.md"
    },
    {
      "key": "diffusion-unet-vs-dit-xl2",
      "a": {
        "key": "diffusion-unet",
        "name": "Diffusion UNet",
        "layers": 17,
        "params": 6687044
      },
      "b": {
        "key": "dit-xl2",
        "name": "DiT-XL/2",
        "layers": 201,
        "params": 670684064
      },
      "question": "Convolutional against transformer backbones for diffusion.",
      "headline": "DiT-XL/2 has 664M more parameters than Diffusion UNet: 199 layers added, 14 removed, 4 changed.",
      "paramDelta": 663997020,
      "layerDelta": 184,
      "added": 199,
      "removed": 14,
      "changed": 4,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/diffusion-unet-vs-dit-xl2.html",
      "markdown": "https://neurarch.com/diff/diffusion-unet-vs-dit-xl2.md"
    },
    {
      "key": "vit-b16-vs-dit-xl2",
      "a": {
        "key": "vit-b16",
        "name": "ViT-B/16",
        "layers": 11,
        "params": 8449000
      },
      "b": {
        "key": "dit-xl2",
        "name": "DiT-XL/2",
        "layers": 201,
        "params": 670684064
      },
      "question": "A vision transformer against a diffusion transformer.",
      "headline": "DiT-XL/2 has 662M more parameters than ViT-B/16: 192 layers added, 1 removed, 9 changed.",
      "paramDelta": 662235064,
      "layerDelta": 190,
      "added": 192,
      "removed": 1,
      "changed": 9,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/vit-b16-vs-dit-xl2.html",
      "markdown": "https://neurarch.com/diff/vit-b16-vs-dit-xl2.md"
    },
    {
      "key": "simple-rnn-vs-mamba-block",
      "a": {
        "key": "simple-rnn",
        "name": "Simple RNN",
        "layers": 2,
        "params": 1100298
      },
      "b": {
        "key": "mamba-block",
        "name": "Mamba SSM Block",
        "layers": 16,
        "params": 168271976
      },
      "question": "The recurrent layer everyone started with against its modern replacement.",
      "headline": "Mamba SSM Block has 167M more parameters than Simple RNN: 15 layers added, 1 removed, 2 changed.",
      "paramDelta": 167171678,
      "layerDelta": 14,
      "added": 15,
      "removed": 1,
      "changed": 2,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/simple-rnn-vs-mamba-block.html",
      "markdown": "https://neurarch.com/diff/simple-rnn-vs-mamba-block.md"
    },
    {
      "key": "cnn-lstm-1d-vs-patch-tst",
      "a": {
        "key": "cnn-lstm-1d",
        "name": "1D CNN + LSTM",
        "layers": 12,
        "params": 311749
      },
      "b": {
        "key": "patch-tst",
        "name": "PatchTST",
        "layers": 17,
        "params": 395396
      },
      "question": "Convolution-plus-recurrence against patched attention for time series.",
      "headline": "PatchTST has 84K more parameters than 1D CNN + LSTM: 14 layers added, 9 removed, 4 changed.",
      "paramDelta": 83647,
      "layerDelta": 5,
      "added": 14,
      "removed": 9,
      "changed": 4,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/cnn-lstm-1d-vs-patch-tst.html",
      "markdown": "https://neurarch.com/diff/cnn-lstm-1d-vs-patch-tst.md"
    },
    {
      "key": "eegnet-vs-eeg-conformer",
      "a": {
        "key": "eegnet",
        "name": "EEGNet",
        "layers": 14,
        "params": 2724
      },
      "b": {
        "key": "eeg-conformer",
        "name": "EEG Conformer",
        "layers": 21,
        "params": 78564
      },
      "question": "A compact EEG convnet against a convolution-plus-transformer one.",
      "headline": "EEG Conformer has 76K more parameters than EEGNet: 12 layers added, 5 removed, 6 changed.",
      "paramDelta": 75840,
      "layerDelta": 7,
      "added": 12,
      "removed": 5,
      "changed": 6,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/eegnet-vs-eeg-conformer.html",
      "markdown": "https://neurarch.com/diff/eegnet-vs-eeg-conformer.md"
    },
    {
      "key": "whisper-small-vs-t5-small",
      "a": {
        "key": "whisper-small",
        "name": "Whisper Small",
        "layers": 15,
        "params": 46986649
      },
      "b": {
        "key": "t5-small",
        "name": "T5 Small",
        "layers": 20,
        "params": 56734592
      },
      "question": "Speech-to-text against text-to-text, both encoder-decoder.",
      "headline": "T5 Small has 9.7M more parameters than Whisper Small: 14 layers added, 9 removed, 8 changed.",
      "paramDelta": 9747943,
      "layerDelta": 5,
      "added": 14,
      "removed": 9,
      "changed": 8,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/whisper-small-vs-t5-small.html",
      "markdown": "https://neurarch.com/diff/whisper-small-vs-t5-small.md"
    },
    {
      "key": "ncf-vs-neumf",
      "a": {
        "key": "ncf",
        "name": "Neural Collaborative Filtering",
        "layers": 9,
        "params": 35206273
      },
      "b": {
        "key": "neumf",
        "name": "Neural Collaborative Filtering",
        "layers": 13,
        "params": 105610401
      },
      "question": "Neural collaborative filtering against its generalised-matrix-factorisation hybrid.",
      "headline": "Neural Collaborative Filtering has 70M more parameters than Neural Collaborative Filtering: 4 layers added, 5 changed.",
      "paramDelta": 70404128,
      "layerDelta": 4,
      "added": 4,
      "removed": 0,
      "changed": 5,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/ncf-vs-neumf.html",
      "markdown": "https://neurarch.com/diff/ncf-vs-neumf.md"
    },
    {
      "key": "wide-and-deep-vs-ncf",
      "a": {
        "key": "wide-and-deep",
        "name": "Wide & Deep",
        "layers": 10,
        "params": 3652882
      },
      "b": {
        "key": "ncf",
        "name": "Neural Collaborative Filtering",
        "layers": 9,
        "params": 35206273
      },
      "question": "The memorisation-plus-generalisation recommender against a purely neural one.",
      "headline": "Neural Collaborative Filtering has 32M more parameters than Wide & Deep: 2 layers added, 3 removed, 5 changed.",
      "paramDelta": 31553391,
      "layerDelta": -1,
      "added": 2,
      "removed": 3,
      "changed": 5,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/wide-and-deep-vs-ncf.html",
      "markdown": "https://neurarch.com/diff/wide-and-deep-vs-ncf.md"
    },
    {
      "key": "lightgcn-vs-graph-sage-rec",
      "a": {
        "key": "lightgcn",
        "name": "LightGCN",
        "layers": 11,
        "params": 70424960
      },
      "b": {
        "key": "graph-sage-rec",
        "name": "GraphSAGE Recommender",
        "layers": 8,
        "params": 263040
      },
      "question": "A stripped graph convolution against a sampled one.",
      "headline": "GraphSAGE Recommender has 70M fewer parameters than LightGCN: 8 layers added, 12 removed, 1 changed.",
      "paramDelta": -70161920,
      "layerDelta": -3,
      "added": 8,
      "removed": 12,
      "changed": 1,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/lightgcn-vs-graph-sage-rec.html",
      "markdown": "https://neurarch.com/diff/lightgcn-vs-graph-sage-rec.md"
    },
    {
      "key": "two-tower-vs-ncf",
      "a": {
        "key": "two-tower",
        "name": "Two-Tower",
        "layers": 10,
        "params": 70433152
      },
      "b": {
        "key": "ncf",
        "name": "Neural Collaborative Filtering",
        "layers": 9,
        "params": 35206273
      },
      "question": "Retrieval by dot product against a learned interaction.",
      "headline": "Neural Collaborative Filtering has 35M fewer parameters than Two-Tower: 3 layers added, 4 removed, 5 changed.",
      "paramDelta": -35226879,
      "layerDelta": -1,
      "added": 3,
      "removed": 4,
      "changed": 5,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/two-tower-vs-ncf.html",
      "markdown": "https://neurarch.com/diff/two-tower-vs-ncf.md"
    },
    {
      "key": "dlrm-vs-wide-and-deep",
      "a": {
        "key": "dlrm",
        "name": "DLRM",
        "layers": 11,
        "params": 64349633
      },
      "b": {
        "key": "wide-and-deep",
        "name": "Wide & Deep",
        "layers": 10,
        "params": 3652882
      },
      "question": "Two production ranking models with different interaction layers.",
      "headline": "Wide & Deep has 61M fewer parameters than DLRM: 3 layers added, 4 removed, 6 changed.",
      "paramDelta": -60696751,
      "layerDelta": -1,
      "added": 3,
      "removed": 4,
      "changed": 6,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/dlrm-vs-wide-and-deep.html",
      "markdown": "https://neurarch.com/diff/dlrm-vs-wide-and-deep.md"
    },
    {
      "key": "bst-vs-sli-rec",
      "a": {
        "key": "bst",
        "name": "Behavior Sequence Transformer",
        "layers": 16,
        "params": 128756259
      },
      "b": {
        "key": "sli-rec",
        "name": "SLi-Rec",
        "layers": 18,
        "params": 64049986
      },
      "question": "Two sequential recommenders with different notions of time.",
      "headline": "SLi-Rec has 65M fewer parameters than Behavior Sequence Transformer: 6 layers added, 5 removed, 10 changed.",
      "paramDelta": -64706273,
      "layerDelta": 2,
      "added": 6,
      "removed": 5,
      "changed": 10,
      "gpuFlips": [],
      "page": "https://neurarch.com/diff/bst-vs-sli-rec.html",
      "markdown": "https://neurarch.com/diff/bst-vs-sli-rec.md"
    },
    {
      "key": "llava-1.5-7b-vs-clip-vit-b32",
      "a": {
        "key": "llava-1.5-7b",
        "name": "LLaVA-1.5",
        "layers": 226,
        "params": 7062339840
      },
      "b": {
        "key": "clip-vit-b32",
        "name": "CLIP ViT-B/32",
        "layers": 37,
        "params": 151198976
      },
      "question": "A vision-language model against the encoder it is built on.",
      "headline": "CLIP ViT-B/32 has 6.9B fewer parameters than LLaVA-1.5: 7 layers added, 196 removed, 32 changed.",
      "paramDelta": -6911140864,
      "layerDelta": -189,
      "added": 7,
      "removed": 196,
      "changed": 32,
      "gpuFlips": [
        {
          "gpu": "T4",
          "a": false,
          "b": true
        },
        {
          "gpu": "A100_40",
          "a": false,
          "b": true
        },
        {
          "gpu": "H100_80",
          "a": false,
          "b": true
        }
      ],
      "page": "https://neurarch.com/diff/llava-1.5-7b-vs-clip-vit-b32.html",
      "markdown": "https://neurarch.com/diff/llava-1.5-7b-vs-clip-vit-b32.md"
    }
  ]
}