Latest in AI
Nemotron 3.5 Super VL: Open Weights, Hardware Requirements and Speed Claims
Nvidia's new 120B vision-language model targets specialized agents. Its deployment recipes explain what the launch's efficiency claims mean in practice.
Nvidia released Nemotron 3.5 Super VL on October 15, 2026, with downloadable weights and hosted access. It combines visual understanding with reasoning and tool use for agents that work across documents, software and business systems.
For teams evaluating it, the hardware configuration and serving recipe deserve as much attention as the benchmark headline.
What is available now
Nvidia's announcement lists Super VL on Hugging Face, ModelScope, OpenRouter and NVIDIA NIM. Nemotron 3.5 Bolt 74B-A9B also became available, fulfilling the date given in Nvidia's earlier RTX Spark software announcement. The smaller Bolt E4B remains scheduled for October 23.
Super VL is the larger multimodal release. Its BF16 model card specifies 120B total parameters, 12B active parameters and support for up to one million context tokens. It accepts text, images and video, and returns text.
The BF16 checkpoint is aimed at evaluation and customization. Its listed deployment configurations include four H100 GPUs or two GB200 GPUs.
The smaller deployment has its own recipe
The NVFP4 model card lists a single DGX Spark, one H200, two H100s or one GB200 as deployment options. Its DGX Spark quick start uses one concurrent sequence and a separate LiLiCorr draft model for speculative decoding.
That configuration detail matters. A working single-user agent and a server supporting several simultaneous users are different capacity tests. A team should benchmark the context lengths, document sizes and concurrency it actually needs before buying hardware around the model.
The weights use OpenMDW-1.1. Nvidia's NIM documentation distinguishes that model license from the separate software terms governing the NIM container.
What the speed comparisons measure
Nvidia's technical post reports up to five times higher throughput than other open models in its size category. Its chart uses 100,000 input tokens, shows more than 600 output tokens per second on the Baseten endpoint and marks a comparison with Inkling Small.
The post separately reports four times faster task completion in a comparison involving 10,000 coding tasks on B200 GPUs. Throughput and completion time measure different things: an agent can generate tokens quickly and still spend too many steps reaching a usable result. These are reported results for particular comparisons, rather than a universal speedup on every machine.
For a useful pilot, give the model representative documents and the same tools as the existing agent. Record whether it finishes correctly, how often someone intervenes and what each accepted result costs. Include difficult cases, such as a misleading chart or a tool returning an incomplete answer.
That test will show whether Super VL's combination of visual input and tool use reduces work for the team, and whether its serving configuration fits the workload.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain AI products through clear technical content and launch stories.