Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
One Model for Vision, Reasoning, and Tool Use
On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, positioning it as an open vision-language model for always-on enterprise agents. It has 120 billion total parameters and activates 12 billion per token. It accepts text, images, document pages, and video, then produces text for reasoning, tool use, and code generation.
The release matters to technical leaders not because replacing separate vision and language models automatically makes an architecture simpler, but because it brings work often split across those models and an agent runtime into one model and inference stack. Integration boundaries may shrink, yet model weights, visual inputs, long context, and multi-turn state still have to work together in deployment. The complexity moves rather than disappears.
Sparse Activation Does Not Mean Lightweight Deployment
Super VL combines Mamba-2 and Transformer layers in its backbone. Visual content passes through a C-RADIO vision encoder and connector before reaching the language backbone. The model card lists 512 routed experts in each MoE layer. Sparse activation lets each token use only part of the model, providing substantial capacity without computing every parameter for every token.
But 12 billion active parameters does not mean that only 12 billion parameters need to be loaded. The total model size still shapes storage and deployment requirements. NVIDIA lists the NVFP4 checkpoint for one DGX Spark, one H200, two H100s, or one GB200, and also publishes a BF16 checkpoint. Nor is NVFP4 uniformly four-bit: the supplied material says the vision encoder, connector, Mamba convolutions, and embeddings remain in BF16, while routers use FP32, with other components using different precisions depending on hardware. Deployment decisions therefore need to account for the full checkpoint, memory, and inference path—not just active parameter count.
Speed Comes from the Whole Inference Path
NVIDIA combines several optimizations. Native multi-token prediction proposes candidate tokens that are then verified in one forward pass. LiLiCorr is a speculative decoding drafter for DGX Spark, which NVIDIA says can raise generation speed by up to 2.3 times. Prefix caching reuses context across agent turns, including attention caches and Mamba states. These techniques target decoding delays and repeated work on prior inputs, and their value depends on how often a task calls the model and how much context can be reused.
NVIDIA’s team also reports more than 600 output tokens per second on Artificial Analysis agentic tasks through a Baseten endpoint, and claims up to five times the single-GPU throughput of open peers. Those figures describe particular serving and evaluation conditions, not a constant that transfers to arbitrary hardware. For engineering teams, the question is whether the optimizations improve their end-to-end requests, including vision encoding, tool wait time, cache hits, and concurrency—not just peak token generation.
The Score Distribution Matters More Than the Top Score
The NVFP4 scores on the model card show strong but uneven performance: 89.90 on GPQA-Diamond, 71.07 on SWE-bench Verified, 60.80 on Terminal-Bench 2.1, 77.5 on BrowseComp, 62.82 on the document task MMLongBench-Doc, and 52.3 on OSWorld-Verified. At the other end, tau3-bench banking scores just 17.78. These results are not a universal proof of agent capability. They show that performance still varies substantially by task type.
Against NVIDIA’s own Nemotron 3 Ultra, Super VL scores 63.44% on Hermes Bench versus 56.33%, and 87.05% on LangChain evaluations versus 84.3%. The supplied material also says Ultra has about 4.6 times as many total parameters. This supports a limited conclusion: on these particular agent evaluations, Super VL’s smaller total parameter count did not prevent it from scoring higher. It does not establish that it will perform better across production tasks, nor does it remove differences in harnesses, tool permissions, or task distributions.
Open Weights Still Require Task and License Review
NVIDIA also released open datasets, training recipes, and reinforcement-learning environments in NeMo Gym, with weights under the OpenMDW-1.1 license. This gives teams a way to study the model and parts of its training process rather than only calling a hosted API. But “open” does not remove the need for review: deployments still require checking license terms, data governance, and the surrounding toolchain.
Capability boundaries also depend on input and runtime conditions. The model card describes video input as MP4 clips of up to two minutes, and reasoning can be toggled through the chat template. The release materials mention Bolt edge models, but Bolt E4B was scheduled for October 23, 2026, so it should not be treated as an available deployment option on the October 15 release date. A sound evaluation is to reproduce key tasks using your own documents, visual inputs, and tools, then calculate the full serving cost. If the workload is chiefly constrained by external tools, persistent state, or domain-specific reliability, more context and faster decoding may not solve the problem.