NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
What changed
NVIDIA launched Nemotron 3.5 Super VL, a multimodal agent model with 120 billion parameters. It uses a mixture-of-experts architecture, but only activates 12 billion parameters at a time for efficiency. The model supports 1 million token context length, enabling processing of very long input sequences. It runs with NVFP4 weights optimized to operate on a single DGX Spark system. NVIDIA also introduced two smaller Bolt edge models based on the same architecture for on-device or low-latency scenarios.
Why builders should care
Nemotron 3.5 Super VL’s long context capability opens new use cases that require extended memory, such as complex document analysis, multi-turn dialogues, or enriched multimodal understanding. The active parameter sparsity means you get a massive underlying model without needing full compute for every query. That lowers inference costs while retaining scale benefits. The inclusion of edge Bolt models means builders can integrate the same advanced architecture in both cloud and edge deployments, improving responsiveness and privacy.
The practical takeaway
For developers working on multimodal AI agents, Nemotron 3.5 Super VL makes it feasible to handle larger, more complex inputs without sprawling infrastructure. Its design is tailored to balance scale with efficiency, reducing the barrier to deploying high-capacity models. Enterprises can also consider the Bolt models for scenarios where latency or connectivity is a bottleneck. This release pressures existing AI stacks to match longer context and multimodal integration while driving expectations for cost-effective scaling.
What to watch next
Focus on performance benchmarks comparing real-world workflows using the full Nemotron 3.5 Super VL versus other large-scale multimodal models. Watch for adoption signals around edge Bolt models in industries where data locality matters. It will also be important to see how quickly the open community extends or customizes Nemotron, and if the NVFP4 weight format gains traction as a new standard for deploying large sparse models efficiently.
AI Quick Briefs Editorial Desk