{
  "site": "https://neurarch.com",
  "description": "Ready-to-open neural network architectures. Each entry links a graph JSON an agent can fetch, edit, and submit to POST /api/v1/check.",
  "count": 36,
  "verify": "https://www.neurarch.com/api/v1/check",
  "mcp": "https://github.com/neurarch-ai/neurarch-mcp",
  "architectures": [
    {
      "key": "eegnet",
      "name": "EEGNet",
      "category": "Biosignal",
      "description": "Compact CNN for EEG/BCI — depthwise + separable convs make it 10× lighter than standard CNNs (Lawhern 2018)",
      "whenToUse": "Pick when channel count is low (4–8) and labelled trials are scarce (<400). Strong default for motor imagery / P300 BCI on consumer headsets.",
      "layers": 16,
      "params": 2724,
      "paramsHuman": "2.7K",
      "inputShape": [
        1,
        22,
        1000
      ],
      "outputShape": [
        4
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/eegnet.html",
      "markdown": "https://neurarch.com/a/eegnet.md",
      "graph": "https://neurarch.com/templates/eegnet/model.json",
      "openInApp": "https://neurarch.com/?template=eegnet"
    },
    {
      "key": "eeg-conformer",
      "name": "EEG Conformer",
      "category": "Biosignal",
      "description": "Conv stem + Transformer encoder — SOTA for high-channel motor imagery EEG (Song 2023)",
      "whenToUse": "Pick when you have ≥16 channels and ~400+ trials per subject. Best published accuracy on BCI IV-2a/2b; expect tricky regularization.",
      "layers": 23,
      "params": 78564,
      "paramsHuman": "78.6K",
      "inputShape": [
        1,
        22,
        1000
      ],
      "outputShape": [
        4
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        }
      ],
      "page": "https://neurarch.com/a/eeg-conformer.html",
      "markdown": "https://neurarch.com/a/eeg-conformer.md",
      "graph": "https://neurarch.com/templates/eeg-conformer/model.json",
      "openInApp": "https://neurarch.com/?template=eeg-conformer"
    },
    {
      "key": "patch-tst",
      "name": "PatchTST",
      "category": "Time-series",
      "description": "Channel-independent patching + Transformer for multivariate time-series (Nie 2023)",
      "whenToUse": "Pick for multivariate time-series with long windows where channels can be treated independently. Same backbone serves classification and forecasting.",
      "layers": 19,
      "params": 395396,
      "paramsHuman": "395.4K",
      "inputShape": [
        1,
        22,
        1000
      ],
      "outputShape": [
        4
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/patch-tst.html",
      "markdown": "https://neurarch.com/a/patch-tst.md",
      "graph": "https://neurarch.com/templates/patch-tst/model.json",
      "openInApp": "https://neurarch.com/?template=patch-tst"
    },
    {
      "key": "cnn-lstm-1d",
      "name": "1D CNN + LSTM",
      "category": "Time-series",
      "description": "Conv1D + LSTM baseline for ECG/PPG/IMU and other long-form physio signals",
      "whenToUse": "Pick for long-form 1D physiological signals (ECG, PPG, IMU) where local morphology + temporal context both matter. Solid baseline before reaching for transformers.",
      "layers": 14,
      "params": 311749,
      "paramsHuman": "311.7K",
      "inputShape": [
        12,
        5000
      ],
      "outputShape": [
        5
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/cnn-lstm-1d.html",
      "markdown": "https://neurarch.com/a/cnn-lstm-1d.md",
      "graph": "https://neurarch.com/templates/cnn-lstm-1d/model.json",
      "openInApp": "https://neurarch.com/?template=cnn-lstm-1d"
    },
    {
      "key": "simple-cnn",
      "name": "Simple CNN",
      "category": "Computer Vision",
      "description": "Simple Convolutional Neural Network for image classification",
      "whenToUse": "Pick as a fast baseline for small images (≤64px, e.g. CIFAR). Trains in minutes, easy to debug — use before reaching for ResNet.",
      "layers": 9,
      "params": 804554,
      "paramsHuman": "804.6K",
      "inputShape": [
        1,
        28,
        28
      ],
      "outputShape": [
        10
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "deep-no-norm",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/simple-cnn.html",
      "markdown": "https://neurarch.com/a/simple-cnn.md",
      "graph": "https://neurarch.com/templates/simple-cnn/model.json",
      "openInApp": "https://neurarch.com/?template=simple-cnn"
    },
    {
      "key": "resnet-block",
      "name": "ResNet Block",
      "category": "Computer Vision",
      "description": "ResNet residual block with skip connections",
      "whenToUse": "Pick when you need a depth-friendly CV backbone. Stack 2–4 blocks for CIFAR, or use as the building block of ResNet-18/50 for ImageNet-scale.",
      "layers": 9,
      "params": 73984,
      "paramsHuman": "74.0K",
      "inputShape": [
        64,
        32,
        32
      ],
      "outputShape": [
        64,
        32,
        32
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/resnet-block.html",
      "markdown": "https://neurarch.com/a/resnet-block.md",
      "graph": "https://neurarch.com/templates/resnet-block/model.json",
      "openInApp": "https://neurarch.com/?template=resnet-block"
    },
    {
      "key": "unet",
      "name": "U-Net",
      "category": "Computer Vision",
      "description": "Encoder-decoder with skip connections — Ronneberger et al. 2015. The standard for biomedical and small-data image segmentation.",
      "whenToUse": "Pick for image segmentation when training data is limited (<10k images). Skip connections preserve fine spatial detail that pure encoder-decoders lose. Default for medical imaging, satellite, and any pixel-level binary mask task.",
      "layers": 24,
      "params": 720705,
      "paramsHuman": "720.7K",
      "inputShape": [
        3,
        256,
        256
      ],
      "outputShape": [
        1,
        256,
        768
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "deep-no-residual",
          "severity": "warn"
        }
      ],
      "page": "https://neurarch.com/a/unet.html",
      "markdown": "https://neurarch.com/a/unet.md",
      "graph": "https://neurarch.com/templates/unet/model.json",
      "openInApp": "https://neurarch.com/?template=unet"
    },
    {
      "key": "vit-b16",
      "name": "ViT-B/16",
      "category": "Computer Vision",
      "description": "Vision Transformer — patch embedding stem + 1 encoder block (768D, 12 heads)",
      "whenToUse": "Pick for 224px+ images when pretrained weights are available, or when dataset is large enough (>1M images) to train from scratch.",
      "layers": 13,
      "params": 8449000,
      "paramsHuman": "8.45M",
      "inputShape": [
        3,
        224,
        224
      ],
      "outputShape": [
        196,
        1000
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "dropout-before-bn",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/vit-b16.html",
      "markdown": "https://neurarch.com/a/vit-b16.md",
      "graph": "https://neurarch.com/templates/vit-b16/model.json",
      "openInApp": "https://neurarch.com/?template=vit-b16"
    },
    {
      "key": "transformer-block",
      "name": "Transformer Block",
      "category": "NLP/LLM",
      "description": "Transformer encoder block (simplified)",
      "whenToUse": "Pick as a generic encoder building block for sequence models when you don't need the modern LLM stack (RoPE/GQA/RMSNorm). Good teaching baseline.",
      "layers": 8,
      "params": 7087872,
      "paramsHuman": "7.09M",
      "inputShape": [
        512,
        768
      ],
      "outputShape": [
        512,
        768
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "bn-at-output",
          "severity": "warn"
        },
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        }
      ],
      "page": "https://neurarch.com/a/transformer-block.html",
      "markdown": "https://neurarch.com/a/transformer-block.md",
      "graph": "https://neurarch.com/templates/transformer-block/model.json",
      "openInApp": "https://neurarch.com/?template=transformer-block"
    },
    {
      "key": "bert-base",
      "name": "BERT Base",
      "category": "NLP/LLM",
      "description": "BERT-Base encoder — bidirectional MHA (12 heads, 768D), no causal mask, 30K vocab",
      "whenToUse": "Pick for text classification / NLU with limited labels — pretrained encoder + small head fine-tunes reliably. 512-token cap.",
      "layers": 11,
      "params": 31120896,
      "paramsHuman": "31.12M",
      "inputShape": [
        1,
        512
      ],
      "outputShape": [
        1,
        512,
        768
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/bert-base.html",
      "markdown": "https://neurarch.com/a/bert-base.md",
      "graph": "https://neurarch.com/templates/bert-base/model.json",
      "openInApp": "https://neurarch.com/?template=bert-base"
    },
    {
      "key": "gpt2",
      "name": "GPT-2",
      "category": "NLP/LLM",
      "description": "GPT-2 Small — causal transformer block (768D, 12 heads, 4× FFN)",
      "whenToUse": "Pick for single-GPU language modeling experiments and as a teaching reference for the canonical decoder-only stack. Modern LLMs prefer LLaMA-3 / Phi-3 blocks.",
      "layers": 12,
      "params": 84331345,
      "paramsHuman": "84.33M",
      "inputShape": [
        1,
        1024
      ],
      "outputShape": [
        1,
        1024,
        50257
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/gpt2.html",
      "markdown": "https://neurarch.com/a/gpt2.md",
      "graph": "https://neurarch.com/templates/gpt2/model.json",
      "openInApp": "https://neurarch.com/?template=gpt2"
    },
    {
      "key": "llama3-block",
      "name": "LLaMA-3 Block",
      "category": "NLP/LLM",
      "description": "LLaMA-3 decoder block — GQA (32H/8KV), SwiGLU FFN, RMSNorm, residual streams",
      "whenToUse": "Pick when building a modern dense LLM (7B–70B class). Strongest open recipe per param; GQA cuts KV cache, RoPE handles long context.",
      "layers": 10,
      "params": 702554112,
      "paramsHuman": "702.55M",
      "inputShape": [
        1,
        2048
      ],
      "outputShape": [
        1,
        2048,
        4096
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/llama3-block.html",
      "markdown": "https://neurarch.com/a/llama3-block.md",
      "graph": "https://neurarch.com/templates/llama3-block/model.json",
      "openInApp": "https://neurarch.com/?template=llama3-block"
    },
    {
      "key": "mixtral-block",
      "name": "Mixtral MoE Block",
      "category": "NLP/LLM",
      "description": "Mixtral decoder block — GQA + Sparse MoE (8 experts, top-2) + RMSNorm + RoPE",
      "whenToUse": "Pick when you have MoE training infra and want best quality per active param. Routing instability and memory cost (still scales with total params) are the trade-off.",
      "layers": 9,
      "params": 1451270144,
      "paramsHuman": "1.45B",
      "inputShape": [
        1,
        4096,
        4096
      ],
      "outputShape": [
        1,
        4096,
        4096
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "moe-no-aux-loss",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/mixtral-block.html",
      "markdown": "https://neurarch.com/a/mixtral-block.md",
      "graph": "https://neurarch.com/templates/mixtral-block/model.json",
      "openInApp": "https://neurarch.com/?template=mixtral-block"
    },
    {
      "key": "t5-small",
      "name": "T5 Small",
      "category": "NLP/LLM",
      "description": "T5 encoder-decoder — bidirectional encoder + masked decoder with cross-attention (512D, 8 heads)",
      "whenToUse": "Pick for seq2seq tasks (summarization, translation, QA) where you need both bidirectional understanding and generation in one model.",
      "layers": 23,
      "params": 56734592,
      "paramsHuman": "56.73M",
      "inputShape": [
        1,
        512
      ],
      "outputShape": [
        1,
        128,
        32128
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        }
      ],
      "page": "https://neurarch.com/a/t5-small.html",
      "markdown": "https://neurarch.com/a/t5-small.md",
      "graph": "https://neurarch.com/templates/t5-small/model.json",
      "openInApp": "https://neurarch.com/?template=t5-small"
    },
    {
      "key": "whisper-small",
      "name": "Whisper Small",
      "category": "Audio",
      "description": "Whisper speech encoder-decoder — conv1d audio stem + transformer encoder/decoder (384D)",
      "whenToUse": "Pick for ASR or as a pretrained audio encoder — drop the decoder + add a head for audio classification (UrbanSound, ESC-50).",
      "layers": 18,
      "params": 46986649,
      "paramsHuman": "46.99M",
      "inputShape": [
        1,
        80,
        3000
      ],
      "outputShape": [
        1,
        448,
        51865
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/whisper-small.html",
      "markdown": "https://neurarch.com/a/whisper-small.md",
      "graph": "https://neurarch.com/templates/whisper-small/model.json",
      "openInApp": "https://neurarch.com/?template=whisper-small"
    },
    {
      "key": "simple-rnn",
      "name": "Simple RNN",
      "category": "NLP",
      "description": "Simple Recurrent Neural Network for sequence processing",
      "whenToUse": "Pick as a teaching reference. Real workloads should reach for LSTM/GRU or transformer — vanilla RNN suffers from vanishing gradients on anything beyond ~50 steps.",
      "layers": 4,
      "params": 1100298,
      "paramsHuman": "1.10M",
      "inputShape": [
        128,
        300
      ],
      "outputShape": [
        128,
        10
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/simple-rnn.html",
      "markdown": "https://neurarch.com/a/simple-rnn.md",
      "graph": "https://neurarch.com/templates/simple-rnn/model.json",
      "openInApp": "https://neurarch.com/?template=simple-rnn"
    },
    {
      "key": "mamba-block",
      "name": "Mamba SSM Block",
      "category": "NLP/LLM",
      "description": "Mamba State Space Model — selective SSM + causal conv gating, no attention (O(T) complexity)",
      "whenToUse": "Pick for very long sequences where attention's O(T²) cost is the bottleneck (DNA, audio, long-context LM). Tuning is trickier than transformers.",
      "layers": 18,
      "params": 168271976,
      "paramsHuman": "168.27M",
      "inputShape": [
        1,
        1024
      ],
      "outputShape": [
        1,
        1024,
        50280
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/mamba-block.html",
      "markdown": "https://neurarch.com/a/mamba-block.md",
      "graph": "https://neurarch.com/templates/mamba-block/model.json",
      "openInApp": "https://neurarch.com/?template=mamba-block"
    },
    {
      "key": "phi3-mini",
      "name": "Phi-3 Mini Block",
      "category": "NLP/LLM",
      "description": "Phi-3 Mini 3.8B decoder block — full MHA (32H, 3072D), SwiGLU FFN (8192), RMSNorm, RoPE",
      "whenToUse": "Pick for compact LLMs (≤4B params) when you want modern ingredients (RoPE, SwiGLU) without the full LLaMA-3 footprint.",
      "layers": 12,
      "params": 310288704,
      "paramsHuman": "310.29M",
      "inputShape": [
        1,
        2048
      ],
      "outputShape": [
        1,
        2048,
        32064
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/phi3-mini.html",
      "markdown": "https://neurarch.com/a/phi3-mini.md",
      "graph": "https://neurarch.com/templates/phi3-mini/model.json",
      "openInApp": "https://neurarch.com/?template=phi3-mini"
    },
    {
      "key": "two-tower",
      "name": "Two-Tower",
      "category": "Recommendation",
      "description": "User+Item dual encoder for retrieval — embeddings → MLP per side → dot product score",
      "whenToUse": "Pick for retrieval at scale (billions of items) where item embeddings can be precomputed and indexed. Not suitable for re-ranking — no cross-features.",
      "layers": 12,
      "params": 70433152,
      "paramsHuman": "70.43M",
      "inputShape": [
        1
      ],
      "outputShape": [
        1,
        64
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "deep-no-norm",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/two-tower.html",
      "markdown": "https://neurarch.com/a/two-tower.md",
      "graph": "https://neurarch.com/templates/two-tower/model.json",
      "openInApp": "https://neurarch.com/?template=two-tower"
    },
    {
      "key": "wide-and-deep",
      "name": "Wide & Deep",
      "category": "Recommendation",
      "description": "Memorization (wide linear) + generalization (deep MLP) joint trained — Cheng et al. 2016",
      "whenToUse": "Pick as a simpler, more interpretable alternative to DLRM for CTR. Linear part captures memorized rules, deep part generalizes.",
      "layers": 13,
      "params": 3652882,
      "paramsHuman": "3.65M",
      "inputShape": [
        10000
      ],
      "outputShape": [
        1
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "output-activation",
          "severity": "info"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "deep-no-norm",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/wide-and-deep.html",
      "markdown": "https://neurarch.com/a/wide-and-deep.md",
      "graph": "https://neurarch.com/templates/wide-and-deep/model.json",
      "openInApp": "https://neurarch.com/?template=wide-and-deep"
    },
    {
      "key": "dlrm",
      "name": "DLRM",
      "category": "Recommendation",
      "description": "Meta's Deep Learning Recommendation Model — bottom MLP for dense, embedding for sparse, feature interaction, top MLP",
      "whenToUse": "Pick when you have substantial dense + sparse features and need a production-validated CTR baseline. Bare-bones — needs feature engineering.",
      "layers": 14,
      "params": 32347553,
      "paramsHuman": "32.35M",
      "inputShape": [
        13
      ],
      "outputShape": [
        32,
        1
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "output-activation",
          "severity": "info"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "deep-no-norm",
          "severity": "info"
        },
        {
          "rule": "init-activation-mismatch",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/dlrm.html",
      "markdown": "https://neurarch.com/a/dlrm.md",
      "graph": "https://neurarch.com/templates/dlrm/model.json",
      "openInApp": "https://neurarch.com/?template=dlrm"
    },
    {
      "key": "neumf",
      "name": "Neural Collaborative Filtering",
      "category": "Recommendation",
      "description": "He et al. 2017 NeuMF — GMF + MLP fused for recommendation",
      "whenToUse": "Pick for pure user/item collaborative filtering when you have no side features. Fused GMF+MLP variant; concat-only variant is `ncf`.",
      "layers": 16,
      "params": 105610401,
      "paramsHuman": "105.61M",
      "inputShape": [
        1
      ],
      "outputShape": [
        1,
        1
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "output-activation",
          "severity": "info"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "deep-no-norm",
          "severity": "info"
        },
        {
          "rule": "init-activation-mismatch",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/neumf.html",
      "markdown": "https://neurarch.com/a/neumf.md",
      "graph": "https://neurarch.com/templates/neumf/model.json",
      "openInApp": "https://neurarch.com/?template=neumf"
    },
    {
      "key": "graph-sage-rec",
      "name": "GraphSAGE Recommender",
      "category": "Recommendation",
      "description": "Inductive node embeddings via neighbor sampling + aggregation — for graph-based recommenders (PinSage style)",
      "whenToUse": "Pick when your recsys has a rich item-item or user-item graph and cold-start items must generalize via neighbors (PinSage-style production setup).",
      "layers": 10,
      "params": 263040,
      "paramsHuman": "263.0K",
      "inputShape": [
        128
      ],
      "outputShape": [
        128,
        64
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "bn-at-output",
          "severity": "warn"
        }
      ],
      "page": "https://neurarch.com/a/graph-sage-rec.html",
      "markdown": "https://neurarch.com/a/graph-sage-rec.md",
      "graph": "https://neurarch.com/templates/graph-sage-rec/model.json",
      "openInApp": "https://neurarch.com/?template=graph-sage-rec"
    },
    {
      "key": "ncf",
      "name": "Neural Collaborative Filtering",
      "category": "Recommendation",
      "description": "He et al. 2017 NCF — user/item embeddings → concat → MLP → score (concat-then-MLP variant)",
      "whenToUse": "Pick as the simplest deep CF baseline before reaching for fused (NeuMF) or graph-based variants. No feature engineering required.",
      "layers": 12,
      "params": 35206273,
      "paramsHuman": "35.21M",
      "inputShape": [
        1
      ],
      "outputShape": [
        1,
        1
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "output-activation",
          "severity": "info"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "deep-no-norm",
          "severity": "info"
        },
        {
          "rule": "init-activation-mismatch",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/ncf.html",
      "markdown": "https://neurarch.com/a/ncf.md",
      "graph": "https://neurarch.com/templates/ncf/model.json",
      "openInApp": "https://neurarch.com/?template=ncf"
    },
    {
      "key": "lightgcn",
      "name": "LightGCN",
      "category": "Recommendation",
      "description": "He et al. 2020 — user/item embeddings propagated through 3 light graph-conv layers, layer combination via mean (no transforms, no nonlinearities)",
      "whenToUse": "Pick for collaborative filtering with implicit feedback (clicks, plays). Strips graph-conv to its essentials — often beats heavier GCN variants on rec benchmarks.",
      "layers": 14,
      "params": 70424960,
      "paramsHuman": "70.42M",
      "inputShape": [
        1
      ],
      "outputShape": [
        64
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "deep-no-norm",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/lightgcn.html",
      "markdown": "https://neurarch.com/a/lightgcn.md",
      "graph": "https://neurarch.com/templates/lightgcn/model.json",
      "openInApp": "https://neurarch.com/?template=lightgcn"
    },
    {
      "key": "bst",
      "name": "Behavior Sequence Transformer",
      "category": "Recommendation",
      "description": "Alibaba 2019 BST — user behavior sequence + target → Transformer encoder + concat with user features → MLP → CTR",
      "whenToUse": "Pick for sequential CTR when user behavior history matters (session intent). Production-deployed at Alibaba scale; needs sequence-aware features.",
      "layers": 20,
      "params": 128756259,
      "paramsHuman": "128.76M",
      "inputShape": [
        50
      ],
      "outputShape": [
        1
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "output-activation",
          "severity": "info"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "deep-no-norm",
          "severity": "info"
        },
        {
          "rule": "consecutive-linear-no-activation",
          "severity": "info"
        },
        {
          "rule": "init-activation-mismatch",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/bst.html",
      "markdown": "https://neurarch.com/a/bst.md",
      "graph": "https://neurarch.com/templates/bst/model.json",
      "openInApp": "https://neurarch.com/?template=bst"
    },
    {
      "key": "sli-rec",
      "name": "SLi-Rec",
      "category": "Recommendation",
      "description": "Yu et al. 2019 — Short and Long-term Interest Recommender. Time-LSTM + ASVD attention fused via gate",
      "whenToUse": "Pick when both short-session intent and long-term preferences matter and you want each modelled separately, then fused via a learned gate.",
      "layers": 20,
      "params": 64049986,
      "paramsHuman": "64.05M",
      "inputShape": [
        50
      ],
      "outputShape": [
        1
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "output-activation",
          "severity": "info"
        },
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "vanishing-gradient",
          "severity": "info"
        },
        {
          "rule": "deep-no-norm",
          "severity": "info"
        },
        {
          "rule": "consecutive-linear-no-activation",
          "severity": "info"
        },
        {
          "rule": "consecutive-linear-no-activation",
          "severity": "info"
        },
        {
          "rule": "init-activation-mismatch",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/sli-rec.html",
      "markdown": "https://neurarch.com/a/sli-rec.md",
      "graph": "https://neurarch.com/templates/sli-rec/model.json",
      "openInApp": "https://neurarch.com/?template=sli-rec"
    },
    {
      "key": "diffusion-unet",
      "name": "Diffusion UNet",
      "category": "Generative",
      "description": "Stable-Diffusion-style noise predictor — latent UNet with cross-attention to a text embedding",
      "whenToUse": "Pick when you want to generate images from a text prompt. The full pipeline also needs a VAE encoder/decoder and a text encoder (e.g. CLIP); this template is the denoiser core.",
      "layers": 19,
      "params": 6687044,
      "paramsHuman": "6.69M",
      "inputShape": [
        4,
        64,
        64
      ],
      "outputShape": [
        4,
        64,
        64
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/diffusion-unet.html",
      "markdown": "https://neurarch.com/a/diffusion-unet.md",
      "graph": "https://neurarch.com/templates/diffusion-unet/model.json",
      "openInApp": "https://neurarch.com/?template=diffusion-unet"
    },
    {
      "key": "deepseek-v3",
      "name": "DeepSeek-V3",
      "category": "NLP/LLM",
      "description": "671B MoE LLM — Multi-head Latent Attention (MLA) compresses the KV cache; fine-grained MoE with shared + routed experts (DeepSeek 2024)",
      "whenToUse": "Study a frontier MoE design: MLA for cheap long-context inference and shared-expert routing. Fold collapses the 61 repeated decoder layers to one block; expand to see the full depth.",
      "layers": 372,
      "params": 666329115904,
      "paramsHuman": "666.33B",
      "inputShape": [
        1,
        163840
      ],
      "outputShape": [
        1,
        163840,
        129280
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "deep-attention-default-init",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/deepseek-v3.html",
      "markdown": "https://neurarch.com/a/deepseek-v3.md",
      "graph": "https://neurarch.com/templates/deepseek-v3/model.json",
      "openInApp": "https://neurarch.com/?template=deepseek-v3"
    },
    {
      "key": "llama-4-scout",
      "name": "Llama-4 Scout",
      "category": "NLP/LLM",
      "description": "109B natively-multimodal MoE LLM — interleaved dense/MoE layers (16 experts) with iRoPE (interleaved no-RoPE) for long context (Meta 2025)",
      "whenToUse": "Reference the newest open MoE decoder: alternating dense and expert-routed blocks. A strong base for studying sparse-activation LLMs.",
      "layers": 434,
      "params": 102547680704,
      "paramsHuman": "102.55B",
      "inputShape": [
        1,
        10485760
      ],
      "outputShape": [
        1,
        10486336,
        202048
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        },
        {
          "rule": "moe-no-aux-loss",
          "severity": "info"
        },
        {
          "rule": "huge-linear-params",
          "severity": "warn"
        },
        {
          "rule": "deep-attention-default-init",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/llama-4-scout.html",
      "markdown": "https://neurarch.com/a/llama-4-scout.md",
      "graph": "https://neurarch.com/templates/llama-4-scout/model.json",
      "openInApp": "https://neurarch.com/?template=llama-4-scout"
    },
    {
      "key": "jamba",
      "name": "Jamba",
      "category": "NLP/LLM",
      "description": "Hybrid SSM-Transformer-MoE — interleaves Mamba, attention, and MoE blocks in one stack (AI21 2024)",
      "whenToUse": "Pick to see three paradigms in one compact graph: state-space, attention, and mixture-of-experts. Good for studying hybrid long-context designs.",
      "layers": 53,
      "params": 13030961152,
      "paramsHuman": "13.03B",
      "inputShape": [
        1,
        4096
      ],
      "outputShape": [
        1,
        4096,
        65536
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        },
        {
          "rule": "moe-no-aux-loss",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/jamba.html",
      "markdown": "https://neurarch.com/a/jamba.md",
      "graph": "https://neurarch.com/templates/jamba/model.json",
      "openInApp": "https://neurarch.com/?template=jamba"
    },
    {
      "key": "qwen3-8b",
      "name": "Qwen3-8B",
      "category": "NLP/LLM",
      "description": "Modern dense decoder LLM — GQA with QK-RMSNorm on the query/key projections for training stability (Alibaba 2025)",
      "whenToUse": "A clean, current dense-decoder reference. Pick when you want a straightforward modern LLM block without MoE routing.",
      "layers": 221,
      "params": 8190878080,
      "paramsHuman": "8.19B",
      "inputShape": [
        1,
        40960
      ],
      "outputShape": [
        1,
        40960,
        151936
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        },
        {
          "rule": "deep-attention-default-init",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/qwen3-8b.html",
      "markdown": "https://neurarch.com/a/qwen3-8b.md",
      "graph": "https://neurarch.com/templates/qwen3-8b/model.json",
      "openInApp": "https://neurarch.com/?template=qwen3-8b"
    },
    {
      "key": "dit-xl2",
      "name": "DiT-XL/2",
      "category": "Generative",
      "description": "Diffusion Transformer — replaces the UNet denoiser with a ViT backbone conditioned on timestep + class via adaLN-Zero (Peebles 2023)",
      "whenToUse": "Pick when you want the modern transformer-based image-generation backbone (used by Sora/SD3-era models) instead of a convolutional UNet denoiser.",
      "layers": 204,
      "params": 670684064,
      "paramsHuman": "670.68M",
      "inputShape": [
        4,
        32,
        32
      ],
      "outputShape": [
        256,
        32
      ],
      "verdict": "pass",
      "findings": [
        {
          "rule": "deep-attention-default-init",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/dit-xl2.html",
      "markdown": "https://neurarch.com/a/dit-xl2.md",
      "graph": "https://neurarch.com/templates/dit-xl2/model.json",
      "openInApp": "https://neurarch.com/?template=dit-xl2"
    },
    {
      "key": "swin-tiny",
      "name": "Swin-Tiny",
      "category": "Computer Vision",
      "description": "Hierarchical vision transformer — shifted-window attention builds a feature pyramid for dense prediction (Liu 2021)",
      "whenToUse": "Pick when you need a ViT that produces multi-scale features (detection, segmentation) rather than a single-scale ViT. Windowed attention keeps compute linear in image size.",
      "layers": 81,
      "params": 28260808,
      "paramsHuman": "28.26M",
      "inputShape": [
        3,
        224,
        224
      ],
      "outputShape": [
        1000
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "attention-no-pe",
          "severity": "warn"
        },
        {
          "rule": "deep-attention-default-init",
          "severity": "info"
        }
      ],
      "page": "https://neurarch.com/a/swin-tiny.html",
      "markdown": "https://neurarch.com/a/swin-tiny.md",
      "graph": "https://neurarch.com/templates/swin-tiny/model.json",
      "openInApp": "https://neurarch.com/?template=swin-tiny"
    },
    {
      "key": "clip-vit-b32",
      "name": "CLIP ViT-B/32",
      "category": "Multimodal",
      "description": "Dual-encoder contrastive model — a ViT image tower and a Transformer text tower projected into a shared embedding space (OpenAI 2021)",
      "whenToUse": "Pick to study cross-modal retrieval / zero-shot classification. Two parallel encoders meet at a contrastive similarity head.",
      "layers": 38,
      "params": 151198976,
      "paramsHuman": "151.20M",
      "inputShape": [
        3,
        224,
        224
      ],
      "outputShape": [
        512
      ],
      "verdict": "pass",
      "findings": [],
      "page": "https://neurarch.com/a/clip-vit-b32.html",
      "markdown": "https://neurarch.com/a/clip-vit-b32.md",
      "graph": "https://neurarch.com/templates/clip-vit-b32/model.json",
      "openInApp": "https://neurarch.com/?template=clip-vit-b32"
    },
    {
      "key": "llava-1.5-7b",
      "name": "LLaVA-1.5",
      "category": "Multimodal",
      "description": "Vision-language model — CLIP image encoder + MLP projector feed visual tokens into a LLaMA decoder (Liu 2023)",
      "whenToUse": "Pick to see the canonical VLM recipe: a frozen vision encoder bridged into an LLM by a small projector. Foundation for image chat / VQA.",
      "layers": 229,
      "params": 7062339840,
      "paramsHuman": "7.06B",
      "inputShape": [
        3,
        336,
        336
      ],
      "outputShape": [
        1,
        2624,
        32000
      ],
      "verdict": "warn",
      "findings": [
        {
          "rule": "full-mha-serving-cost",
          "severity": "info"
        },
        {
          "rule": "deep-attention-default-init",
          "severity": "info"
        },
        {
          "rule": "kv-cache-context-budget",
          "severity": "warn"
        }
      ],
      "page": "https://neurarch.com/a/llava-1.5-7b.html",
      "markdown": "https://neurarch.com/a/llava-1.5-7b.md",
      "graph": "https://neurarch.com/templates/llava-1.5-7b/model.json",
      "openInApp": "https://neurarch.com/?template=llava-1.5-7b"
    }
  ]
}