Nvidia’s Nemotron 3.5 Super VL Runs AI Agents 5x Faster — and the Weights Are Open

Introduction: Speed Is the New Frontier

On October 15, 2026, Nvidia released Nemotron 3.5 Super VL, an open 120-billion-parameter model the company says delivers “the fastest task completion and highest accuracy for specialized always-on agents” — 5x faster than the next-best open model in its class. The catch that isn’t a catch: the weights are fully downloadable, under Nvidia’s open MDW license, on day one.

For a year, the open-weight leaderboard has read like a Chinese lab roster — GLM, Kimi, DeepSeek, Qwen trading the top spots. Super VL, built with the multi-lab Nemotron Coalition, is Nvidia’s loudest statement yet that US labs can compete at open weights on the metric that increasingly decides agent deployments: not raw intelligence, but completed tasks per dollar per second. An agent that runs for days or weeks multiplies every inefficiency; a model that answers 5x faster at frontier-adjacent accuracy changes the economics of the whole workload.

What Super VL Actually Is

Nemotron 3.5 Super VL is a mixture-of-experts model with 120 billion total parameters but only 12 billion active per token, paired with a C-RADIO vision encoder and up to 1 million tokens of context. The architecture is a hybrid Mamba-Transformer backbone with multi-token prediction and NVFP4 quantization — the plumbing that lets it hit over 600 output tokens per second on agentic tasks even with 100,000 tokens of input context, according to Nvidia’s measurements on a Baseten endpoint.

The “VL” is the other half of the story: one model handles text, images, documents, and video alongside reasoning, instruction following, and tool use. That matters for real agent pipelines, where a single request — read a contract, check a policy, call a tool, verify against a screenshot — otherwise means shuttling context between a text model and a separate vision-language model, losing state at every hop. Super VL collapses that into one deployment.

It ships with Day 0 support in vLLM, and the weights are already on Hugging Face. In Nvidia’s system-of-models framing, Super VL sits between Nemotron 3.5 Lightning (high-volume text) and Nemotron 3 Ultra (hardest reasoning) as the default execution engine for multimodal agent work.

The Benchmarks: Beating Nvidia’s Own Frontier Model

The most striking numbers aren’t against rivals — they’re against Nvidia’s own flagship. On agent-harness evaluations published today:

A model at roughly one-fifth the active-parameter count outscoring the big sibling on agentic tasks is the clearest evidence yet that 2026’s training recipes reward agent-shaped workloads over general benchmark bulk. On Artificial Analysis testing cited by ZDNet, the speed gap over similarly sized open rivals (InclusionAI’s Ling 3.0 Flash and DeepSeek’s V4.1 Flash came second and third) was decisive. Raw coding tells a more mixed story — 71.1% on SWE-Bench Verified is solid but well behind Claude Opus 5’s 96% — which is exactly why Nvidia positions Super VL as a specialist: the always-on workhorse, not the heavyweight.

The Post-Training Multiplier

Open weights are only as valuable as your ability to customize them, and Nvidia led with its best anecdote: post-training a base Super VL on domain data produced a 44-point gain on medical insurance claims tasks, taking it “far above Claude Opus 5” on that specialized workload. A general model becomes a specialist, and the enterprise keeps the model, the data, and the infrastructure under its own roof — post-training runs through Nvidia’s NeMo toolkit with your own tools, policies, and workflows.

That’s the pitch closed APIs can’t make. It echoes what we saw in Nvidia’s harness research: agentic performance is increasingly engineered around the model, not just inside it — and you can only engineer freely when you can post-train, quantize, and deploy the weights yourself.

Nemotron 3.5 Bolt: Agents on Your Desk

Super VL didn’t arrive alone. Two edge models complete the family:

• Nemotron 3.5 Bolt 74B-A9B 24GB — a 74B MoE with 9B active, squeezed into 24GB via native 2-bit quantization so it runs on an RTX or DGX Spark alongside your other apps. Nvidia claims it’s the #1 open model in its class on PinchBench and #2 open on BrowseComp, targeting local agent harnesses for research, writing, and debugging with zero cloud dependence.
• Nemotron 3.5 Bolt E4B 4GB — a 4GB model for fast conversational agents on GeForce RTX, Jetson, and DGX Spark, arriving October 23. Four gigabytes for a multi-turn conversational agent is the kind of number that puts AI assistants into devices that could never have hosted one.

The Bolt line continues the local-AI momentum we tracked in Nvidia’s RTX SPARK push: privacy-sensitive workloads staying on-device, agents that keep working offline, and no per-token bill for the agent that never sleeps.

The US Open-Weight Surge After the Hugging Face Hack

As ZDNet notes, these releases land amid an open-model renaissance with a specific catalyst: the Hugging Face incident spooked enterprises about supply-chain risk, and interest in self-hosted, auditable weights jumped. Western labs have responded — Nvidia’s Nemotron Coalition joins Reflection AI’s Beam and Mistral’s trillion-parameter open-weight push in a surge that’s finally contesting the Chinese labs’ open-weight lead. For buyers, the shift is concrete: the “open vs. closed” debate is becoming a procurement question about data control and deployment sovereignty, not an ideology.

What This Means for the Tools You Pick

If you’re assembling an agent stack in late 2026, today’s release changes three calculations:

• Price throughput per completed task, not per token. A 5x faster model at slightly lower accuracy often finishes more work per dollar than a slower genius — measure your harness end to end.
• One multimodal model beats two specialists. Collapsing text + vision pipelines into a single Super VL-style model removes context hand-offs, the fragile seam in most agent failures.
• Check the exit ramp. Hosted platforms like Replicate, Together AI, and AWS SageMaker make open weights operational — but the point of open weights is that you can always leave, post-train on your data, and redeploy on your own GPUs.

Frequently Asked Questions