NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B on October 15, 2026, targeting multimodal agents and local agents, respectively. Bolt E4B 4GB is scheduled for October 23, 2026. Sources: NVIDIA
Super VL’s hybrid Mamba-2/Transformer mixture-of-experts architecture has 120 billion total parameters and 12 billion active; it accepts text, images and video and produces text. NVIDIA lists availability through Hugging Face, ModelScope, OpenRouter and NVIDIA NIM, its packaged inference microservice. Downloadable weights support self-hosting, customization and evaluation, while inference services provide access to a running model. Sources: NVIDIA
Bolt’s official 24GB GGUF card specifies IQ2 2-bit quantization with quantization-aware training, a main model, a separate multimodal projector and an optional speculative-decoding drafter. NVIDIA’s architecture-search target is below 24GB for weights and runtime states at batch size one and 128K context. The card also lists 28GB and 32GB deployment recipes, so neither the target nor weight storage alone establishes a workload’s total runtime memory needs. Sources: NVIDIA
NVIDIA announced Bolt support through llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Developers should verify the exact artifact revision, runtime version and hardware configuration with their intended workload before deployment. OpenHands described itself as an NVIDIA launch partner and said it was bringing Super VL into ongoing software-engineering workflows. Sources: NVIDIA, OpenHands
NVIDIA claims up to fivefold higher Super VL throughput than models in its class and ranks Bolt first in its class on PinchBench. However, the Bolt card lists PinchBench scores of 87.82 for Bolt and 88.07 for Qwen3.6-35B-A3B BF16. Comparative conclusions need clear class definitions and matched baselines, workloads, hardware and evaluation methods; OpenHands’ statement provides partner testimony rather than independent validation. Sources: NVIDIA, OpenHands