SGLang added day-0 serving support for NVIDIA Nemotron 3.5 Super VL on October 15, 2026, with deployment instructions for running the model through its serving layer. NVIDIA describes Nemotron 3.5 Super VL as an open multimodal model that accepts text, images, documents, and video and produces text for agent reasoning and tool use. Its mixture-of-experts architecture has 120 billion total parameters and activates 12 billion per token.

The model is positioned as the high-accuracy multimodal execution model in the Nemotron 3 family, alongside Nemotron 3.5 Lightning and Nemotron 3 Ultra. NVIDIA says organizations can adapt it with NeMo for domain-specific terminology, tasks, policies, and tool-use patterns. Suggested workflows include interpreting customer-service screenshots and forms, comparing reports and scanned documents, supporting multi-step work in business software, analyzing video, and conducting web research.

SGLang supports continuous batching, prefix caching, speculative decoding, and OpenAI-compatible APIs for serving the model. The guide describes multi-token prediction (MTP), which proposes several future tokens for verification during decoding, and reports 243.9 tokens per second on a single RTX 6000 PRO. NVIDIA lists deployment targets including DGX Spark, 1x GB10, B200, B300, RTX PRO 6000, H200, and 2x H100.

NVIDIA says the model’s hybrid Mamba-Transformer design, MTP, LatentMoE, and NVFP4 quantization are intended to improve throughput and memory efficiency for long-context multimodal agent tasks. Its figures report up to 5x faster throughput for agentic tasks than other open models in its size category on a single GPU, over 600 TPS on Artificial Analysis Intelligence Index agentic tasks on a Baseten endpoint, and 4x faster task completion than other leading open models of similar size. The guide provides BF16 and NVFP4 weights through Hugging Face and a getting-started cookbook for deployment with SGLang.

SGLang added day-0 serving support for NVIDIA Nemotron 3.5 Super VL on October 15, 2026, with deployment instructions for running the model through its serving layer. NVIDIA describes Nemotron 3.5 Super VL as an open multimodal model that accepts text, images, documents, and video and produces text for agent reasoning and tool use. Its mixture-of-experts architecture has 120 billion total parameters and activates 12 billion per token.

The model is positioned as the high-accuracy multimodal execution model in the Nemotron 3 family, alongside Nemotron 3.5 Lightning and Nemotron 3 Ultra. NVIDIA says organizations can adapt it with NeMo for domain-specific terminology, tasks, policies, and tool-use patterns. Suggested workflows include interpreting customer-service screenshots and forms, comparing reports and scanned documents, supporting multi-step work in business software, analyzing video, and conducting web research.

SGLang supports continuous batching, prefix caching, speculative decoding, and OpenAI-compatible APIs for serving the model. The guide describes multi-token prediction (MTP), which proposes several future tokens for verification during decoding, and reports 243.9 tokens per second on a single RTX 6000 PRO. NVIDIA lists deployment targets including DGX Spark, 1x GB10, B200, B300, RTX PRO 6000, H200, and 2x H100.

NVIDIA says the model’s hybrid Mamba-Transformer design, MTP, LatentMoE, and NVFP4 quantization are intended to improve throughput and memory efficiency for long-context multimodal agent tasks. Its figures report up to 5x faster throughput for agentic tasks than other open models in its size category on a single GPU, over 600 TPS on Artificial Analysis Intelligence Index agentic tasks on a Baseten endpoint, and 4x faster task completion than other leading open models of similar size. The guide provides BF16 and NVFP4 weights through Hugging Face and a getting-started cookbook for deployment with SGLang.