NVIDIA Nemotron 3.5 Bolt 74B-A9B

Open-weight VLM, weights created on Hugging Face 2026-10-09 (nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16); the card's Release Date line says 2026-10-15, which is when the card text was finalized. Sits just outside the strict sub-8B-active bar (9B active) but inside the consumer-runnable class the card is built around: the companion NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB checkpoint is a native 2-bit (IQ2) quantization that the card says fits in 24GB and is the recommended inference artifact (llama.cpp GGUF and vLLM). This record describes the BF16 parent, not the quant. Card: 74B total / 9B active, Mamba-2-Transformer hybrid LatentMoE, C-RADIO v4-H vision encoder (vit_hidden_size 1280), context 256K, inputs text/image/video. Compressed from Nemotron 3.5 Super VL 120B-A12B with Star Elastic (arXiv 2605.07182). config.json llm_config (model_type nemotron_h_puzzle): hidden 4096, 88 blocks = 40 mamba + 40 moe + 8 attention (indices 7, 16, 25, 36, 47, 58, 69, 78), attention heads 32 / kv heads 2 / head_dim 128 so gqa_ratio 16, vocab 131072, max_position_embeddings 262144, sliding_window null. MoE is not uniform: every MoE block routes num_experts_per_tok 9, but n_routed_experts varies (12 blocks at 240, 6 at 304, 9 at 368, 10 at 432, 3 at 496) and moe_intermediate_size varies (1536/2048/2304/2560). Plus n_shared_experts 1 at moe_shared_expert_intermediate_size 5376 and moe_latent_size 1024, so experts_per_token 9 is the routed count and the shared expert is extra. Mamba-2: mamba_num_heads 128, mamba_head_dim 64, ssm_state_size 128, conv_kernel 4. KV (bf16, attention blocks only): 8 layers x 2 kv heads x 128 x 2 x 2 = 8192 B/token = 8 KiB/token. The Mamba recurrent state is the other memory term and is not in that number. flops_per_token_active = 2 x 9e9 = 1.8e10 is the schema approximation from the vendor's 9B-active figure; hybrid Mamba plus variable-width experts mean it is not exact. Benchmarks on the card (Terminal-Bench 2.1 53.93 BF16 / 50.98 IQ2, GPQA Diamond 88.83 / 86.43, HLE no-tools 28.73 / 25.30) are vendor_claim. License OpenMDW-1.1, commercial use allowed per the card.