NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 model card and config.json

Primary source for the Nemotron 3.5 Bolt record. HF API createdAt 2026-10-09; the card's own Release Date line says Hugging Face 10/15/2026 (the day the card was finished; weights were already up). Card states 74B total / 9B active, Mamba-2 + Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder, context up to 256K, OpenMDW-1.1, compressed from Nemotron 3.5 Super VL 120B-A12B via Star Elastic (arXiv 2605.07182). config.json llm_config (model_type nemotron_h_puzzle, fetched 2026-10-16) confirms hidden_size 4096, 88 blocks, num_attention_heads 32, num_key_value_heads 2, head_dim 128, vocab_size 131072, max_position_embeddings 262144, num_experts 512, n_shared_experts 1, moe_shared_expert_intermediate_size 5376, moe_latent_size 1024, mamba_num_heads 128, mamba_head_dim 64, ssm_state_size 128. layers_block_type counts: 40 mamba, 40 moe, 8 attention (indices 7, 16, 25, 36, 47, 58, 69, 78). Every MoE block sets num_experts_per_tok 9; n_routed_experts varies by block (240/304/368/432/496). Benchmark table on the card is a vendor claim. Companion 24GB IQ2 checkpoint is a quantization of this model, not a separate architecture.