NVIDIA Launches Nemotron 3.5 Super VL Multimodal AI

NVIDIA has introduced Nemotron 3.5 Super VL, an open 120B-parameter mixture-of-experts model that activates 12B parameters per token. Built to streamline agentic workflows, the system consolidates text, image, document, and video understanding alongside instruction following and tool use into one model, bypassing the need for separate vision-language models. On a single GPU, it delivers up to 5x higher throughput compared to size-class competitors and over 600 TPS on Baseten endpoints evaluating Artificial Analysis Intelligence Index tasks, finishing workloads 4x faster overall.

Distilled from Nemotron 3 Ultra using SkewReverse loss and multi-teacher on-policy distillation, the model surpasses its predecessor on Hermes Bench (63.44% versus 56.33%) and LangChain behavioral evaluations (87.05% versus 84.3%), while outperforming Nemotron 3 Super on SWE-bench and Terminal-Bench. On the CharXiv chart-reasoning benchmark, it registered 80.8% accuracy—outstripping Nemotron 3 Nano Omni’s 63%—and hit 84% on Python code execution while topping open models on the OCRBench v2 leaderboard. Early adopters report drastic performance gains: Scale AI saw a 40x cost drop and 5x latency reduction for Skydance Corporation's credit processing, Ramp expects 50% higher throughput and 10x lower latency for receipt parsing, and AI21 with Alice boosted threat analysis accuracy by 41%.

To train the architecture, NVIDIA leveraged tools like NeMo Data Designer alongside contributions from coalition partners such as Perplexity and Prime Intellect. A continued pretraining phase incorporated 10B Indic tokens, leading to a broader release with IIT Madras and AI4Bharat containing 2.8T tokens across 22 Indian languages. Developers can further customize the model using NVIDIA NeMo for supervised fine-tuning, LoRA, and reinforcement learning with NeMo Gym, supported by 40 vision-language datasets published on Hugging Face.

For inference efficiency, the model uses LiLiCorr speculative decoding to boost token generation speed by up to 2.3x on DGX Spark hardware. Post-training NVFP4 quantization with quantization-aware distillation enables 4-bit execution on Blackwell architecture and 4-bit weights with 16-bit activations elsewhere, maintaining KV-cache accuracy. Combined with multi-token prediction and Mamba state preservation, deployment via NVIDIA Dynamo on 16 B200 GPUs running at 50 tokens per second per user can serve 8x more users with an 85% drop in first-token latency, alongside packaging through NVIDIA NIM and access via OpenRouter.