NVIDIA Releases Open Nemotron 3.5 Super VL for Agents

NVIDIA has released Nemotron 3.5 Super VL, an open-weight vision-language model with 120 billion total parameters, alongside two smaller "Bolt" models built to run locally. The launch, dated October 15, 2026, matters because it pairs image, document, and video understanding with a design aimed at long-running software agents, and because the weights are available to developers rather than locked behind a single hosted service.

What NVIDIA released

According to Interesting Engineering, which cites NVIDIA's own announcement, the flagship Nemotron 3.5 Super VL has 120 billion parameters in total, of which 12 billion are active for any given token. That ratio points to a mixture-of-experts design, in which only part of the network runs for each step. NVIDIA describes the model as handling reasoning, tool use, and the understanding of text, images, documents, and video, with multistep and long-running agent tasks as the target workload. LetsDataScience adds that the model is meant for agents that work across charts, web pages, and documents.

Two smaller models round out the family. Nemotron 3.5 Bolt 74B-A9B uses native 2-bit quantization so that it fits in 24GB of memory, which puts it within reach of a single high-end consumer or workstation GPU. NVIDIA positions it for local coding assistance and delegated web research. Nemotron 3.5 Bolt E4B is a 4GB model for fast conversational agents, including voice assistants and interactive characters, and for physical AI systems on NVIDIA Jetson hardware. Both outlets report that the E4B model is scheduled for October 23; the 74B-A9B model is available now.

Claims, and who is making them

Most of the headline numbers are NVIDIA's. The company says Super VL completes tasks five times faster than the nearest competitor in its class and approaches the accuracy of NVIDIA's largest reasoning model while using about one-fifth as many parameters. It also says the Bolt 74B-A9B ranks first among open models in its class on PinchBench and second among open models on BrowseComp, although the coverage we reviewed gives no scores for either result. These are company claims, and readers should treat them as such until independent evaluations appear.

Some independent measurement already exists. LetsDataScience reports that Artificial Analysis lists Super VL at a 26.5 score on its Intelligence Index and at about 630 output tokens per second with 100,000-token prompts, though that throughput figure comes from a pre-release endpoint. The same source lists the Bolt 74B-A9B at 74 billion total and 9 billion active parameters, with a 21.4 Intelligence Index score. A pre-release endpoint is not the same as the production services most teams will use, so the throughput number is best read as an indication rather than a guarantee. Both articles also note that benchmark results do not replace testing on a team's own workflows, such as document accuracy, tool-calling reliability, latency, and cost.

Training and availability

Interesting Engineering reports that NVIDIA trained the model on roughly 26 trillion tokens of pretraining data, followed by about 3.8 million post-training samples and reinforcement learning across dozens of environments. That recipe, a very large pretraining run followed by a comparatively small post-training set and then reinforcement learning, mirrors the pattern that has become standard for agent-oriented models. The relevant point for builders is that the reinforcement learning stage is where tool use and multistep behaviour are typically sharpened.

The model ships with open weights and supporting training materials. It is available on Hugging Face, OpenRouter, and ModelScope, and through NVIDIA NIM, and it can be customised through NVIDIA NeMo and Python-based pipelines. The coverage does not state the licence terms, which is an important detail for any company considering commercial deployment. Anyone planning to build on the weights should read the licence on the model card before committing.

Early adopters

NVIDIA named several early users, according to Interesting Engineering. Scale AI is using the model for a credit-finance agent built for a Skydance vendor and for a separate M&A due diligence agent for EY. AfterQuery is applying it to credit agreement analysis for Hebbia. AI21 and Alice are testing resilience against cyberattacks, and Mercor evaluated it on coding, using the APEX-SWE benchmark. The pattern is notable: most of the named workloads involve long, document-heavy professional tasks where a model must read charts, tables, and scanned pages and then act on what it finds, which is exactly where a vision-language model with tool use is meant to help.

These are descriptions of pilots and evaluations supplied by NVIDIA, not independent audits, and the coverage gives no outcome data for any of them.

Why it matters for multimodal AI

The release reflects a broader shift in how multimodal models are positioned. Earlier vision-language models were often framed as tools for describing or captioning images. Here the framing is operational: a model that can look at a document, a web page, or a video frame and decide what to do next. That makes the vision component an input to an agent loop rather than an end product, and it raises the bar on reliability, since an agent that misreads a chart can propagate the error through every later step.

The efficiency angle is equally important. Running only 12 billion parameters per token is what allows a 120-billion-parameter model to claim high throughput. For companies paying for inference at scale, the cost of an agent that runs for minutes or hours is dominated by how many tokens it generates and how quickly, so active-parameter count matters more than the headline total. The Bolt models push the same logic to the edge, trading some capability for the ability to run on a single machine without sending data to a hosted API, which can matter for privacy-sensitive work.

What to watch next

Several open questions will determine whether the launch has lasting impact. First, independent evaluations on document understanding, video reasoning, and tool-use benchmarks will show whether the speed and accuracy claims hold up outside NVIDIA's own tests. Second, the licence terms will decide how freely companies can fine-tune and deploy the weights. Third, the October 23 arrival of the Bolt E4B model will show whether a 4GB conversational model is good enough for the voice and robotics use cases NVIDIA has described.

It is also worth watching how competing open-weight multimodal models respond. NVIDIA's strategy of releasing capable open models supports demand for its hardware and software stack, and each release raises the baseline rivals must meet. For developers, the practical step is straightforward: try the models on a representative slice of real work, measure accuracy and cost against current choices, and treat the vendor's comparisons as a starting hypothesis rather than a conclusion.