NVIDIA has launched Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model engineered to run enterprise AI agents continuously across background tasks. Built using a Mixture-of-Experts architecture, the model activates only 12 billion parameters per token and was distilled from NVIDIA’s Nemotron 3 Ultra in partnership with the Nemotron Coalition, which includes Mistral AI, Perplexity, and Black Forest Labs.

Super VL outperforms its predecessor on coding benchmarks such as SWE-bench and matches larger models on tool use, visual understanding, and multi-step workflows. Hardware optimizations enable the model to achieve up to 5x higher throughput on a single GPU compared to similar open models, utilizing 4-bit NVFP4 quantization, speculative draft token verification, and computation caching for recurring system prompts.

NVIDIA is positioning the model, alongside smaller local hardware variants in the Bolt series, for enterprise customers looking to post-train and host proprietary agents on local servers or DGX Spark hardware using the NeMo framework.

Why it matters

  • Enables enterprises to run agentic visual and code workflows on lower compute footprints via 4-bit NVFP4 quantization.

  • Reduces latency for repetitive agent system prompts using integrated draft components and hardware computation caching.

  • Offers a customizable open-weights alternative to closed proprietary APIs for regulatory-heavy industries like finance and insurance.

Source: the-decoder.com