Nemotron 3.5 Super VL: Nvidia's Open Model for Always-On Agents
Agentic AINvidia says its new open model matches the accuracy of its largest reasoning model on agent tasks with about one-fifth of the parameters, and it is pitching that efficiency at agents that run for days or weeks without stopping. The Decoder’s report on Nemotron 3.5 Super VL describes a 120-billion-parameter multimodal model built for enterprise agents that read documents, tables, charts and video. For teams running agents over sensitive data, the more unusual part is what ships with it: datasets and training recipes, not only weights.
What Happened
Super VL uses a mixture-of-experts design with 12 billion of its 120 billion parameters active per token, according to Nvidia’s technical blog post. Nvidia distilled it from Nemotron 3 Ultra and says it matches Ultra on instruction following, tool calling, multi-step business processes and image analysis. In agent tests from Hermes and LangChain, Nvidia says it scores higher than Ultra. On the chart benchmark CharXiv it scores 80.8%, rising to 84% when it can run Python, by Nvidia’s figures.
Two smaller Bolt models ship alongside it. Bolt 74B-A9B 24GB runs agent tasks on local machines, using 2-bit quantization to fit in 24 GB of memory. A 4 GB Bolt E4B for voice assistants is due later this month.
Why It Matters
Always-on agents resend the same instructions, tool descriptions and history with every request. Super VL caches those recurring computations, and Nvidia’s Dynamo inference software routes a session’s requests to the GPU that already holds its cache. On 16 B200 GPUs, Nvidia says this lets the model serve eight times as many concurrent users, with time to first token down 85%. A built-in draft component proposes tokens that the main model verifies in one pass, and on DGX Spark Nvidia says its LiLiCorr drafting method speeds generation by up to 2.3x over running without drafts.
Openness goes further than usual. Nvidia is releasing base, post-trained and 4-bit weights on Hugging Face, plus training datasets, recipes and reinforcement learning environments. The Decoder notes the catch: everything is tuned for Nvidia hardware, so getting the most from Nemotron means buying Nvidia’s chips.
What This Means for Builders
Nvidia’s pitch is fine-tuning on private data and running the model on your own servers, free of a provider’s terms of service. Early adopters it names are document-heavy. Ramp turns receipt photos into structured data, EY uses it for acquisition due diligence, and Hebbia analyzes credit agreements. Scale AI adapted it for a vendor-credits agent at Skydance, with costs expected to drop to one-fortieth. Those are expectations reported by Nvidia, not published results.
The local Bolt model is less accessible than it sounds. The Decoder points out that 24 GB of memory is a high bar that only expensive hardware meets, so for now the local agent targets developers and owners of powerful workstations. Stack Archive tracked the strategy behind this lineup in Nvidia’s $26B Open-Source Bet That Could Reshape Agentic AI.
The So What
Enterprise teams building document agents on regulated data should shortlist Super VL against their current hosted model, starting with one document-heavy workflow and their own accuracy and cost measurements, since every benchmark in the launch is Nvidia’s. Teams without Nvidia GPUs should weigh the hardware dependence before investing in fine-tuning, because the caching and 4-bit gains are tuned for Nvidia’s stack.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →