NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning

NVIDIA has announced Nemotron 3.5 Super VL, an open, customizable multimodal model aimed at always-on enterprise agents that handle long, multi-step workflows. The company describes it as built for tasks where a single request can involve reading a document, checking information across sources, following company policy and calling several tools.
The model is a 120B-parameter mixture-of-experts (MoE) architecture that activates 12B parameters per token. This is the core efficiency trade-off of MoE designs: the full model holds a large amount of capacity, but only a fraction of it is used for any given token, which keeps per-token compute far below what a dense 120B model would require. NVIDIA says the model delivers up to 5x higher throughput for agentic tasks than other open models in its size category on a single GPU, and reports over 600 tokens per second on Artificial Analysis Intelligence Index agentic tasks on the Baseten endpoint.
Capability-wise, Nemotron 3.5 Super VL combines reasoning, instruction following and tool use with text, image, document and video understanding in one model. NVIDIA's pitch is deployment consolidation: teams handling both textual and visual work no longer need separate text and vision-language models, or the task context passing that would otherwise sit between them. Named workflows include coding (testing and validating code, analyzing errors, recommending fixes), web research (searching sites, connecting information, verifying answers), document intelligence (extracting from tables, charts and page layouts), cybersecurity (identifying vulnerabilities and triaging threats) and video understanding for search, content review and incident analysis.
NVIDIA positions the model inside a broader portfolio rather than as a standalone answer. It says smaller models such as Nemotron 3.5 Lightning are intended for high-volume text tasks, while frontier models such as Nemotron 3 Ultra tackle complex planning and orchestration, leaving Nemotron 3.5 Super VL for specialized multistep work that needs reasoning, tool use and mixed text and visual understanding together.
Customization runs through NVIDIA NeMo. Developers can post-train the model with their own data, tools, policies and workflows, using NeMo AutoModel for supervised fine-tuning (SFT) and low-rank adaptation (LoRA), while reinforcement learning is described as improving performance in agentic environments that reflect organizational workflows. NVIDIA is also releasing open datasets on Hugging Face, including agentic tool-use and conversational data, finance reasoning and tool-calling datasets, instruction-following and chat data, long-context transduction and translation data, fairness data and 45 vision-language datasets.
NVIDIA cites several production examples. Scale AI post-trained the model for automated vendor credit processing at Skydance Corporation, with inference cost expected to fall 40x and latency 5x. Ramp is converting receipt images into structured JSON containing merchant information, dates and payment data, and expects up to 50% higher throughput and 10x lower latency. AI21 post-trained it for Alice's threat intelligence and vulnerability analysis, improving accuracy by 40%.
On visual benchmarks, NVIDIA says the model scored 80.8% on CharXiv, a chart-reasoning benchmark, compared with 63% for Nemotron 3 Nano Omni, and reached 84% with Python code interpretation. It is described as one of the top open models on the OCRBench v2 leaderboard. Evaluations span Hermes, LangChain, OpenClaw, OpenCode, Pi, Codex and Claude, and the model scores higher than Nemotron 3 Ultra on Hermes Bench and LangChain behavioral evaluations. NVIDIA also reports 4x faster task completion through higher throughput and token efficiency, and says an independent EXO Labs evaluation placed it on the Pareto frontier for accuracy and task completion time on DGX Spark.
The training pipeline is layered. The model was distilled from Nemotron 3 Ultra, with a SkewReverse loss used during distillation. Post-training uses multi-teacher on-policy distillation (MOPD), in which the model generates responses and specialized teachers score them for video reasoning and tracking, vision tool use and alignment. Continued pretraining added 10B Indic tokens drawn from a 12-language subset of Nemotron-pretraining-Indic-v1 and natural Indic text, and NVIDIA is releasing the full roughly 2.8T-token, 22-language Indic dataset developed with IIT Madras and its AI4Bharat lab. Figure 5 in the post covers Indic benchmark accuracy.
Several inference techniques target the repetition inherent in agent turns. LiLiCorr is an open speculative decoding method that jointly evaluates token candidates across an entire draft to capture dependencies independent token prediction misses; a lightweight network computes compatibility scores in one pass, leaving a low-cost sequential final selection. NVIDIA calls it the recommended drafter on DGX Spark, delivering up to 2.3x faster token generation than decoding without speculation. Multi-token prediction adds an internal drafter that proposes several tokens for the main model to verify in one forward pass. Prefix caching reuses computation from shared inputs such as system instructions and tool definitions, and the serving stack preserves both attention caches and Mamba recurrent states.
On precision, the model uses NVIDIA Model Optimizer for post-training quantization and quantization-aware distillation (QAD). The NVFP4 recipe uses 4-bit weights and activations on Blackwell; the same checkpoint runs on non-Blackwell platforms with 4-bit weights and 16-bit activations. QAD trains the quantized model to match the original BF16 outputs and can also store the key-value cache in NVFP4 instead of FP8, adding no inference overhead because it happens during training.
Deployment is offered through NVIDIA NIM, which packages optimized engines, validated configurations and standard APIs, and NVIDIA Dynamo for distributed inference, including KV-cache-aware routing and disaggregated serving. NVIDIA says that on 16 B200 GPUs at 50 tokens per second per user, KV-aware routing with session affinity and KV cache offload serves up to 8x as many users while cutting first-token latency by 85%. The release includes base, post-trained and NVFP4-quantized checkpoints with training recipes, data and reinforcement-learning environments, plus a usage cookbook on GitHub, a technical report, Hugging Face or ModelScope deployment instructions and hosted access via OpenRouter. A long partner list spans customization, serving and cloud providers, including Baseten, Databricks, Fireworks AI, Together AI, CoreWeave, Modal and vLLM, Ollama and SGLang on the inference side.
Why it matters: teams building document-heavy or visual agent workflows get a single open checkpoint that covers reasoning, tool use and multimodal input, with weights they can post-train and host locally or in their own cloud. The claimed throughput and latency gains matter most for always-on agents, where per-task cost and completion time, not peak benchmark scores, determine how much work a deployment can finish. The practical caveats are that many headline numbers come from NVIDIA's own evaluations or partner pilots, and realizing the NVFP4 efficiency on Blackwell, or LiLiCorr's speedup on DGX Spark, depends on adopting the recommended quantization and serving stack rather than running the weights as-is.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Always-on agents are taking on longer, more complex enterprise workflows. A single request can involve reading a document, checking information across sources,...