Nvidia's Nemotron 3.5 Super VL posts 5x agentic speed lead over open rivals

Nvidia released the open-weight Nemotron 3.5 Super VL on Tuesday, claiming the 120-billion-parameter model runs agentic tasks five times faster than the next-fastest open model in its class. The benchmark finding comes from independent evaluator Artificial Analysis, which placed InclusionAI's Ling 3.0 Flash second and DeepSeek's V4.1 Flash third, with a measurable distance separating Super VL from both.

What Nvidia is selling

Nvidia framed Super VL as a tool for "the fastest task completion and highest accuracy for specialized, always-on agents." The company said in its release that the model "helps enterprise agents complete high-volume tasks faster across workflows that span days or weeks," a positioning aimed squarely at deployments that require persistent autonomous execution rather than single-turn chat.

Super VL targets an industry pivot toward agentic AI, the term for systems that plan, call tools, and complete multi-step tasks without continuous human supervision. Artificial Analysis' testing measured throughput on those workflows rather than raw question-answer accuracy.

What the benchmarks show

Nvidia cited a 44 percent post-training gain on medical insurance claims processing, a result that pushed Super VL above Anthropic's Claude Opus 5 on that specific task. The base version of the model did not top proprietary frontier systems on its own, the company acknowledged. The 44-point lift came from domain-specific fine-tuning enterprises can apply on their own infrastructure.

That customization angle is the central pitch for open-weight releases over closed alternatives. Nvidia stated that businesses can post-train the model on proprietary data and run it on their own hardware, a combination the company describes as offering greater personalization and security than closed APIs permit.

How does Super VL compare with named rivals?

• Super VL: 120 billion parameters; fastest in class per Artificial Analysis
• Ling 3.0 Flash (InclusionAI): second in Artificial Analysis ranking
• V4.1 Flash (DeepSeek): third in Artificial Analysis ranking
• Claude Opus 5 (Anthropic): exceeded by Super VL on medical insurance claims after post-training

Two edge models ship alongside Super VL

Nvidia released two smaller models built for local, offline deployment:

• Nemotron 3.5 Bolt 74B-A9B 24GB: designed for local agentic use cases
• Nemotron 3.5 Bolt E4B 4GB: pitched by Nvidia for "fast, conversational agents"

Edge AI refers to models that operate on local hardware without constant cloud connectivity. The two Bolt releases extend Nvidia's open-weight footprint into that category, where parameter count and memory footprint determine whether a model fits on a workstation or a server.

The open-weight race is shifting

The release lands inside a broader reshuffling of the open-weight market, a segment Chinese labs have dominated since DeepSeek's early advances. Three recent Western releases point to a closing geography gap:

• Mistral launched Le Chonk, which posted state-of-the-art results on certain benchmarks
• Reflection AI released Beam last week; Nvidia has invested in Reflection and is in talks to acquire the company
• Nvidia shipped Super VL and the two Bolt models on Tuesday

Open models from Chinese firms such as DeepSeek and Moonshot.ai established the template for cheaper, customizable alternatives to closed frontier models from OpenAI, Anthropic, and Google. Western open-weight efforts have lagged in user adoption until recently.

Why enterprise buyers are reconsidering open weights

Industry adoption of open-weight models has accelerated since a breach disclosed earlier this year on Hugging Face, the dominant open-model hub. The incident challenged the prior assumption that only open models posed security exposure. Closed-model providers faced similar scrutiny after the disclosure.

That shift matters for procurement. Cybersecurity guidance in several regulated sectors already restricts the use of foreign-developed AI software, a posture that has sidelined Chinese open-weight models inside some banks, defense contractors, and government agencies. US-made open models that match Chinese benchmarks on accuracy and speed give those buyers a domestic alternative without surrendering the customization rights open weights provide.

Artificial Analysis' efficiency findings are the immediate evidence. The broader question is whether Super VL sustains its throughput lead once enterprises apply the same post-training recipes to competing models.

What to watch next

Three developments will determine whether the geography gap in open-weight AI stays closed:

• Whether Reflection AI's pending deal with Nvidia closes and what GPU allocation the combined entity secures
• Whether Mistral's Le Chonk holds its benchmark positions against follow-up releases from DeepSeek and Moonshot
• Whether enterprise post-training case studies replicate the 44 percent lift Nvidia reported on medical insurance claims

Nvidia priced Super VL at no cost as an open-weight download, the standard distribution model for the category. The competitive question is not price; it is throughput per parameter, where Artificial Analysis now ranks Super VL first by a factor of five.