74B total / 9B active (LatentMoE)
262,144 tokens (256K)
OpenMDW License Agreement, version 1.1
Open-weight vision-language reasoning and agent model compressed from NVIDIA Nemotron 3.5 Super VL 120B-A12B for local deployment. Bolt combines a hybrid Mamba-2, Transformer, Mixture-of-Experts, and attention language backbone with a C-RADIO v4-H vision encoder and MLP projector, accepting text, images, and video while producing text. NVIDIA used its Star Elastic router-guided architecture-search and knowledge-distillation method to reduce the parent model from 120B total / 12B active parameters to 74B total / 9B active, lowering expert top-K from 22 to 9. The native IQ2 release uses quantization-aware training to pack the model into approximately 24GB, allowing it to run alongside local agent applications; full BF16 and IQ2 GGUF checkpoints are also available. Training included compression and recovery over approximately 80B tokens, vision-adapter recovery, vision-heavy supervised fine-tuning, checkpoint merging, RLHF, and end-to-end GRPO in MCP and software-engineering environments. The model supports switchable thinking and non-thinking modes, tool calling, web browsing, terminal use, document and screen understanding, and English, Spanish, French, German, Italian, Japanese, and programming languages. A companion LiLiCorr draft model uses DFlash to propose seven-token blocks for speculative decoding and reports a 4.30-token average accepted length on SPEED-Bench. NVIDIA reports IQ2 scores of 50.98 on Terminal-Bench 2.1, 86.43 on GPQA Diamond, 66.75 on BrowseComp, 52.20 on MCP-Atlas, and 50.86 on vision WebArena; the BF16 checkpoint scores 53.93, 88.83, 70.00, 55.60, and 50.25 respectively.