{ "count": 19, "license": "CC-BY-SA-4.0", "generated": "2026-04-29T10:37:47.522Z", "items": [ { "id": "qwen2.5-coder-32b", "name": "Qwen2.5-Coder 32B Instruct", "lab": "alibaba", "release_date": "2024-11-12", "license": "Apache-2.0", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "dense", "total_params_b": 32, "active_params_b": 32, "layers": 64, "hidden_size": 5120, "ffn_size": 27648, "num_attention_heads": 40, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 152064, "max_context_length": 131072, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 60500000000, "bytes_per_token": 60500000000 }, { "operator": "attention", "flops_per_token": 3500000000, "bytes_per_token": 3500000000 }, { "operator": "rmsnorm", "flops_per_token": 6553600, "bytes_per_token": 819200 }, { "operator": "rope", "flops_per_token": 327680, "bytes_per_token": 40960 }, { "operator": "silu", "flops_per_token": 121634816, "bytes_per_token": 60817408 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "qwen3.5-397b", "name": "Qwen3.5 397B Reasoning", "lab": "alibaba", "release_date": "2026-03-05", "license": "Apache-2.0", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 397, "active_params_b": 22, "layers": 64, "hidden_size": 5120, "ffn_size": 14336, "num_attention_heads": 40, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 152064, "max_context_length": 131072, "moe": { "num_experts": 128, "top_k": 8, "expert_hidden_size": 1536, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 28185722880, "bytes_per_token": 28185722880 }, { "operator": "attention", "flops_per_token": 9395240960, "bytes_per_token": 14495514624 }, { "operator": "moe-gate", "flops_per_token": 41943040, "bytes_per_token": 8053063680 }, { "operator": "rmsnorm", "flops_per_token": 3276800, "bytes_per_token": 1310720 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "qwen3.6-plus", "name": "Qwen3.6 Plus", "lab": "alibaba", "release_date": "2026-03-25", "license": "Apache-2.0", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 480, "active_params_b": 35, "layers": 64, "hidden_size": 6144, "ffn_size": 16384, "num_attention_heads": 64, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 152064, "max_context_length": 1048576, "moe": { "num_experts": 128, "top_k": 8, "expert_hidden_size": 2048, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 38654705664, "bytes_per_token": 38654705664 }, { "operator": "attention", "flops_per_token": 12884901888, "bytes_per_token": 20132659200 }, { "operator": "moe-gate", "flops_per_token": 50331648, "bytes_per_token": 12884901888 }, { "operator": "rmsnorm", "flops_per_token": 3932160, "bytes_per_token": 1572864 } ], "modalities": [ "text", "vision" ], "weight_format": "bf16" }, { "id": "deepseek-r1", "name": "DeepSeek R1", "lab": "deepseek", "release_date": "2025-01-20", "license": "deepseek-license", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 671, "active_params_b": 37, "layers": 61, "hidden_size": 7168, "ffn_size": 18432, "num_attention_heads": 128, "num_kv_heads": 128, "head_dim": 128, "vocab_size": 129280, "max_context_length": 131072, "moe": { "num_experts": 256, "top_k": 8, "expert_hidden_size": 2048, "shared_experts": 0 }, "attention_type": "mla" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 48356130816, "bytes_per_token": 48356130816 }, { "operator": "attention", "flops_per_token": 16118710272, "bytes_per_token": 32237420544 }, { "operator": "moe-gate", "flops_per_token": 111935488, "bytes_per_token": 14327742464 }, { "operator": "rmsnorm", "flops_per_token": 4372480, "bytes_per_token": 1748992 } ], "modalities": [ "text" ], "weight_format": "bf16", "hf_url": "https://huggingface.co/deepseek-ai/DeepSeek-R1" }, { "id": "deepseek-v4-flash", "name": "DeepSeek V4 Flash", "lab": "deepseek", "release_date": "2026-04-24", "license": "deepseek-license", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 284, "active_params_b": 13, "layers": 32, "hidden_size": 4096, "ffn_size": 14336, "num_attention_heads": 32, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 132000, "max_context_length": 1048576, "moe": { "num_experts": 64, "top_k": 4, "expert_hidden_size": 1408, "shared_experts": 0 }, "attention_type": "csa+hca" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 11274289152, "bytes_per_token": 11274289152 }, { "operator": "attention", "flops_per_token": 3221225472, "bytes_per_token": 4831838208 }, { "operator": "moe-gate", "flops_per_token": 8388608, "bytes_per_token": 1476395008 }, { "operator": "rmsnorm", "flops_per_token": 1310720, "bytes_per_token": 524288 } ], "modalities": [ "text" ], "weight_format": "bf16", "hf_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash" }, { "id": "deepseek-v4-pro", "name": "DeepSeek V4 Pro", "lab": "deepseek", "release_date": "2026-04-24", "license": "deepseek-license", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 1600, "active_params_b": 49, "layers": 64, "hidden_size": 8192, "ffn_size": 24576, "num_attention_heads": 64, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 132000, "max_context_length": 1048576, "moe": { "num_experts": 256, "top_k": 8, "expert_hidden_size": 2048, "shared_experts": 0 }, "attention_type": "csa+hca", "rope_theta": 10000000 }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 4800000000, "bytes_per_token": 18000000 }, { "operator": "attention", "flops_per_token": 1200000000, "bytes_per_token": 4500000 }, { "operator": "moe-gate", "flops_per_token": 10000000, "bytes_per_token": 1000000 }, { "operator": "rmsnorm", "flops_per_token": 5000000, "bytes_per_token": 1000000 } ], "modalities": [ "text" ], "weight_format": "bf16", "hf_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro" }, { "id": "alphafold-3", "name": "AlphaFold 3", "lab": "google", "release_date": "2024-05-08", "license": "creative-commons-by-nc-4.0", "domain": "scientific", "workload_kind": "forward-only-batch", "architecture": { "family": "dense", "total_params_b": 0.21, "active_params_b": 0.21, "layers": 48, "hidden_size": 384, "ffn_size": 1536, "num_attention_heads": 16, "num_kv_heads": 16, "head_dim": 24, "vocab_size": 64, "max_context_length": 5120, "attention_type": "pair-bias-attention" }, "operator_decomposition": [ { "operator": "pair-bias-attention", "flops_per_token": 2400000000, "bytes_per_token": 800000000, "notes": "Pair representation update — quadratic in residue count, dominant cost" }, { "operator": "matmul", "flops_per_token": 1200000000, "bytes_per_token": 400000000, "notes": "PairFormer + Diffusion linear projections" }, { "operator": "triangle-multiplication", "flops_per_token": 600000000, "bytes_per_token": 600000000, "notes": "AF3-specific triangle update — FP32-sensitive numerics" }, { "operator": "rmsnorm", "flops_per_token": 8000000, "bytes_per_token": 4000000 } ], "modalities": [ "text" ], "weight_format": "bf16", "paper_url": "https://www.nature.com/articles/s41586-024-07487-w", "github_url": "https://github.com/google-deepmind/alphafold3", "notes": "AlphaFold 3 (Abramson et al. 2024, Nature). Predicts joint structure of\nproteins, DNA, RNA, ligands, and ions. Uses PairFormer + Diffusion\nModule instead of AlphaFold 2's Evoformer + IPA.\n\nCritical for scientific accelerators: this is NOT a transformer LLM\nworkload. Pair-bias attention is O(N^2) in residue count (typical\nN=2000), making memory-bandwidth dominant for large complexes.\nFP32 precision needed in triangle multiplication for numerical\nstability; FP16/BF16 work for most other ops.\n\nInference is single-shot (forward-only-batch), not autoregressive.\nLatency target: 1-30 minutes per structure depending on size.\n" }, { "id": "gemma-4", "name": "Gemma 4 26B", "lab": "google", "release_date": "2026-02-20", "license": "Gemma-License", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 26, "active_params_b": 4, "layers": 32, "hidden_size": 4096, "ffn_size": 11264, "num_attention_heads": 32, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 256000, "max_context_length": 131072, "moe": { "num_experts": 16, "top_k": 2, "expert_hidden_size": 5632, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 8858370048, "bytes_per_token": 8858370048 }, { "operator": "attention", "flops_per_token": 3221225472, "bytes_per_token": 4831838208 }, { "operator": "moe-gate", "flops_per_token": 2097152, "bytes_per_token": 2952790016 }, { "operator": "rmsnorm", "flops_per_token": 1310720, "bytes_per_token": 524288 } ], "modalities": [ "text", "vision" ], "weight_format": "bf16" }, { "id": "graphcast", "name": "GraphCast", "lab": "google", "release_date": "2023-11-14", "license": "creative-commons-by-nc-sa-4.0", "domain": "scientific", "workload_kind": "graph-iteration", "architecture": { "family": "dense", "total_params_b": 0.0367, "active_params_b": 0.0367, "layers": 16, "hidden_size": 512, "ffn_size": 1024, "num_attention_heads": 8, "num_kv_heads": 8, "head_dim": 64, "vocab_size": 256, "max_context_length": 1038240, "attention_type": "graph-message-passing" }, "operator_decomposition": [ { "operator": "graph-message-passing", "flops_per_token": 280000000, "bytes_per_token": 280000000, "notes": "Mesh GNN message passing — 16 processor steps, edge-bound" }, { "operator": "matmul", "flops_per_token": 180000000, "bytes_per_token": 60000000, "notes": "Encoder/decoder MLPs + per-edge MLPs" }, { "operator": "scatter-gather", "flops_per_token": 40000000, "bytes_per_token": 200000000, "notes": "Grid ↔ icosahedral mesh interpolation — bandwidth-bound" }, { "operator": "layernorm", "flops_per_token": 4000000, "bytes_per_token": 2000000 } ], "modalities": [ "text" ], "weight_format": "fp32", "paper_url": "https://www.science.org/doi/10.1126/science.adi2336", "github_url": "https://github.com/google-deepmind/graphcast", "notes": "GraphCast (Lam et al. 2023, Science). 0.25° / 1° global weather\nforecast at 6-hour intervals, up to 10 days. Outperforms HRES\n(operational numerical weather prediction) on 90% of variables.\n\nCritical scientific-computing workload characteristic: graph-based\niterative inference (40 forward passes for a 10-day forecast) over\n~1M-vertex mesh. NOT autoregressive in the LLM sense — each step\nproduces global state.\n\nMemory-bandwidth-bound on accelerators with HBM (1M vertex × 474\nvariables = ~2 GB activations per step). Triangulated icosahedral\nmesh creates irregular memory access patterns that favor cards with\nlarge cache hierarchies (H100 50 MB L2, MI300X 256 MB Infinity Cache)\nover those with raw HBM bandwidth alone.\n" }, { "id": "llama-3.3-70b", "name": "Llama 3.3 70B Instruct", "lab": "meta", "release_date": "2024-12-06", "license": "Llama-3.3-Community", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "dense", "total_params_b": 70, "active_params_b": 70, "layers": 80, "hidden_size": 8192, "ffn_size": 28672, "num_attention_heads": 64, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 128256, "max_context_length": 131072, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 132000000000, "bytes_per_token": 132000000000 }, { "operator": "attention", "flops_per_token": 7500000000, "bytes_per_token": 7500000000 }, { "operator": "rmsnorm", "flops_per_token": 8400000, "bytes_per_token": 1310720 }, { "operator": "rope", "flops_per_token": 524288, "bytes_per_token": 65536 }, { "operator": "silu", "flops_per_token": 156905472, "bytes_per_token": 78643200 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "llama-4-maverick", "name": "Llama 4 Maverick", "lab": "meta", "release_date": "2025-04-05", "license": "Llama-4-Community", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 400, "active_params_b": 17, "layers": 48, "hidden_size": 5120, "ffn_size": 16384, "num_attention_heads": 40, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 200000, "max_context_length": 1048576, "moe": { "num_experts": 128, "top_k": 1, "expert_hidden_size": 8192, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 24159191040, "bytes_per_token": 24159191040 }, { "operator": "attention", "flops_per_token": 7046430720, "bytes_per_token": 10871635968 }, { "operator": "moe-gate", "flops_per_token": 31457280, "bytes_per_token": 4026531840 }, { "operator": "rmsnorm", "flops_per_token": 2457600, "bytes_per_token": 983040 } ], "modalities": [ "text", "vision", "video" ], "weight_format": "bf16" }, { "id": "llama-4-scout", "name": "Llama 4 Scout", "lab": "meta", "release_date": "2025-04-05", "license": "Llama-4-Community", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 109, "active_params_b": 17, "layers": 48, "hidden_size": 5120, "ffn_size": 16384, "num_attention_heads": 40, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 200000, "max_context_length": 10485760, "moe": { "num_experts": 16, "top_k": 1, "expert_hidden_size": 8192, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 24159191040, "bytes_per_token": 24159191040 }, { "operator": "attention", "flops_per_token": 7046430720, "bytes_per_token": 10871635968 }, { "operator": "moe-gate", "flops_per_token": 3932160, "bytes_per_token": 4026531840 }, { "operator": "rmsnorm", "flops_per_token": 2457600, "bytes_per_token": 983040 } ], "modalities": [ "text", "vision", "video" ], "weight_format": "bf16" }, { "id": "minimax-m2.7", "name": "MiniMax M2.7", "lab": "minimax", "release_date": "2026-04-10", "license": "MIT", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "hybrid", "total_params_b": 456, "active_params_b": 46, "layers": 80, "hidden_size": 6144, "ffn_size": 16384, "num_attention_heads": 64, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 200000, "max_context_length": 4194304, "moe": { "num_experts": 32, "top_k": 2, "expert_hidden_size": 16384, "shared_experts": 0 }, "attention_type": "lightning" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 48318382080, "bytes_per_token": 48318382080 }, { "operator": "attention", "flops_per_token": 16106127360, "bytes_per_token": 25165824000 }, { "operator": "moe-gate", "flops_per_token": 15728640, "bytes_per_token": 32212254720 }, { "operator": "rmsnorm", "flops_per_token": 4915200, "bytes_per_token": 1966080 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "mistral-large-3", "name": "Mistral Large 3", "lab": "mistral", "release_date": "2025-08-14", "license": "MRL-2", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "dense", "total_params_b": 123, "active_params_b": 123, "layers": 88, "hidden_size": 12288, "ffn_size": 28672, "num_attention_heads": 96, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 32768, "max_context_length": 128000, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 232000000000, "bytes_per_token": 232000000000 }, { "operator": "attention", "flops_per_token": 12500000000, "bytes_per_token": 12500000000 }, { "operator": "rmsnorm", "flops_per_token": 14745600, "bytes_per_token": 1966080 }, { "operator": "rope", "flops_per_token": 786432, "bytes_per_token": 98304 }, { "operator": "silu", "flops_per_token": 252706816, "bytes_per_token": 126353408 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "mistral-small-4", "name": "Mistral Small 4", "lab": "mistral", "release_date": "2026-03-16", "license": "Apache-2.0", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 119, "active_params_b": 22, "layers": 40, "hidden_size": 5120, "ffn_size": 14336, "num_attention_heads": 32, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 131072, "max_context_length": 131072, "moe": { "num_experts": 8, "top_k": 2, "expert_hidden_size": 14336, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 17616076800, "bytes_per_token": 17616076800 }, { "operator": "attention", "flops_per_token": 5872025600, "bytes_per_token": 9227468800 }, { "operator": "moe-gate", "flops_per_token": 1638400, "bytes_per_token": 11744051200 }, { "operator": "rmsnorm", "flops_per_token": 2048000, "bytes_per_token": 819200 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "kimi-k2.6", "name": "Kimi K2.6", "lab": "moonshot", "release_date": "2026-04-15", "license": "moonshot-license", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 1000, "active_params_b": 32, "layers": 60, "hidden_size": 7168, "ffn_size": 18432, "num_attention_heads": 64, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 160000, "max_context_length": 262144, "moe": { "num_experts": 384, "top_k": 8, "expert_hidden_size": 1536, "shared_experts": 0 }, "attention_type": "mla" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 47563407360, "bytes_per_token": 47563407360 }, { "operator": "attention", "flops_per_token": 15854469120, "bytes_per_token": 25543311360 }, { "operator": "moe-gate", "flops_per_token": 165150720, "bytes_per_token": 10569646080 }, { "operator": "rmsnorm", "flops_per_token": 4300800, "bytes_per_token": 1720320 } ], "modalities": [ "text", "vision" ], "weight_format": "bf16", "hf_url": "https://huggingface.co/moonshotai/Kimi-K2.6" }, { "id": "gpt-oss", "name": "GPT-OSS", "lab": "openai", "release_date": "2025-08-12", "license": "Apache-2.0", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 120, "active_params_b": 5, "layers": 36, "hidden_size": 2880, "ffn_size": 11520, "num_attention_heads": 32, "num_kv_heads": 8, "head_dim": 90, "vocab_size": 200000, "max_context_length": 131072, "moe": { "num_experts": 128, "top_k": 4, "expert_hidden_size": 2880, "shared_experts": 0 }, "attention_type": "gqa" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 7166361600, "bytes_per_token": 7166361600 }, { "operator": "attention", "flops_per_token": 2043740160, "bytes_per_token": 2813460480 }, { "operator": "moe-gate", "flops_per_token": 13271040, "bytes_per_token": 2388787200 }, { "operator": "rmsnorm", "flops_per_token": 1036800, "bytes_per_token": 414720 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "glm-5-reasoning", "name": "GLM-5 Reasoning", "lab": "zhipu", "release_date": "2026-03-15", "license": "MIT", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "dense", "total_params_b": 32, "active_params_b": 32, "layers": 60, "hidden_size": 5120, "ffn_size": 17920, "num_attention_heads": 40, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 151552, "max_context_length": 131072, "attention_type": "mha" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 33030144000, "bytes_per_token": 33030144000 }, { "operator": "attention", "flops_per_token": 8808038400, "bytes_per_token": 13589544960 }, { "operator": "rmsnorm", "flops_per_token": 3072000, "bytes_per_token": 1228800 } ], "modalities": [ "text" ], "weight_format": "bf16" }, { "id": "glm-5.1", "name": "GLM-5.1", "lab": "zhipu", "release_date": "2026-04-07", "license": "MIT", "domain": "llm", "workload_kind": "autoregressive-decode", "architecture": { "family": "moe", "total_params_b": 754, "active_params_b": 32, "layers": 64, "hidden_size": 6144, "ffn_size": 16384, "num_attention_heads": 48, "num_kv_heads": 8, "head_dim": 128, "vocab_size": 151552, "max_context_length": 131072, "moe": { "num_experts": 192, "top_k": 8, "expert_hidden_size": 1536, "shared_experts": 0 }, "attention_type": "mha" }, "operator_decomposition": [ { "operator": "matmul", "flops_per_token": 38654705664, "bytes_per_token": 38654705664 }, { "operator": "attention", "flops_per_token": 12884901888, "bytes_per_token": 20401094656 }, { "operator": "moe-gate", "flops_per_token": 75497472, "bytes_per_token": 9663676416 }, { "operator": "rmsnorm", "flops_per_token": 3932160, "bytes_per_token": 1572864 } ], "modalities": [ "text" ], "weight_format": "bf16", "hf_url": "https://huggingface.co/THUDM/GLM-5.1" } ] }