NVIDIA releases Nemotron 3.5 Super VL and Bolt models

NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B on October 15. The company said a smaller Bolt E4B model would follow on October 23. The release extends the Nemotron 3.5 family from high-volume agent tasks to multimodal server workflows and local deployment.

The additions broaden NVIDIA’s open-model lineup across several hardware targets. Nemotron 3.5 Super VL accepts text, images and video for reasoning, coding, tool use, document analysis and video understanding. NVIDIA positions the Bolt models for agent workloads on desktops and edge systems.

Super VL has 120 billion total parameters and activates 12 billion for each token, according to NVIDIA’s model card. Its mixture-of-experts architecture routes each input through part of the model rather than using every parameter for every token. NVIDIA lists deployments on one DGX Spark, H200 or GB200 system, or on two H100 GPUs. Some listed configurations support context windows of up to 1 million tokens.

NVIDIA says Super VL can deliver up to five times the generation throughput of another open model in its size or performance class under the company’s stated test conditions. The comparison does not establish that end-to-end workflows finish five times faster. None of the opened sources independently reproduced NVIDIA’s throughput, latency or task-completion claims.

Bolt 74B-A9B has 74 billion total parameters and activates 9 billion at a time. NVIDIA says native 2-bit quantization, which stores model weights with fewer bits, lets the model fit in 24GB of memory. The company positions it for local agents on RTX systems and DGX Spark, with support through llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

NVIDIA positions Bolt E4B for multi-turn conversation and physical-agent applications on GeForce RTX, Jetson, and RTX or DGX Spark systems. Its announcement did not disclose scores or define the comparison class behind the Bolt ranking claims. Bolt E4B was not available at the research cutoff.

The additions extend the family beyond Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model NVIDIA released in August for high-volume agent tasks. Super VL adds image and video inputs for larger deployments, while the Bolt models target machines with tighter memory limits.

NVIDIA released the quantized Super VL checkpoint under its OpenMDW-1.1 license, which the model card says permits commercial and noncommercial use. The company says the broader release includes model checkpoints, training recipes, datasets and reinforcement-learning environments. Super VL is available through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM; Bolt 74B-A9B is available through Hugging Face.