Nvidia is positioning Nemotron 3.5 Super VL as an open model for enterprise AI agents that work continuously, handling tasks across text, images, documents and video. Alongside it, the company is introducing smaller Bolt models designed to bring agent capabilities to local computers and edge devices.
A large model built for repeated work
Nemotron 3.5 Super VL has 120 billion parameters and uses a Mixture-of-Experts architecture. Nvidia says 12 billion parameters are active for each token, allowing the model to handle agent tasks while limiting how much of the model is engaged at a time.
The model is intended for software that follows instructions, calls tools and works through multi-step business processes. It can also analyze images and documents, making it relevant to organizations where agents need to interpret information in different formats as part of ongoing workflows.
Nvidia says Super VL was distilled from Nemotron 3 Ultra. On some agent tasks, the company says it matches Ultra’s accuracy with about one-fifth the parameters. Nvidia reports higher scores than Ultra in tests from Hermes and LangChain, and says Super VL also outperforms Nemotron 3 Super on SWE-bench and Terminal-Bench.
These results are company claims, and the article also describes an independent EXO Labs test on the DGX Spark. Nvidia says the model delivers up to five times the throughput of similarly sized open models on a single GPU, citing the Artificial Analysis leaderboard. In the EXO Labs test, Nvidia says no other model was both faster and more accurate.
Designed to keep computing costs in check
Agents often repeat the same instructions, tool descriptions and conversation history with each request. Super VL can cache computations for these recurring parts, reducing the need to process them again. Nvidia’s Dynamo software can also direct requests from a session to a GPU that already holds its cache.
Nvidia reports that, on 16 B200 GPUs, this setup serves eight times as many concurrent users, while time to first token falls by 85 percent. For agents that may handle large volumes of work over days or weeks, keeping repeated computation down can matter alongside the model’s raw speed.
The company also uses its 4-bit NVFP4 format to compress Super VL. On Blackwell GPUs, the model runs entirely at 4-bit precision; on older cards, weights are stored at 4 bits while calculations use 16-bit precision. Nvidia says sensitive parts keep higher precision and that the compressed model was trained to match the original’s outputs.
A draft component can propose several tokens for the main model to verify together. For DGX Spark, Nvidia recommends LiLiCorr, which evaluates draft tokens together. Nvidia says this approach can make text generation up to 2.3x faster than running without speculative drafts.
Document analysis is central to the vision features
Super VL is designed to work with material that appears in everyday business workflows, including OCR, tables, charts, multi-page documents and video. It can connect a chart to a table elsewhere in the same document and summarize both, without requiring a separate vision model.
On CharXiv, Nvidia reports a score of 80.8 percent, compared with 63 percent for Nemotron 3 Nano Omni. When the model can also run Python code, the score rises to 84 percent. Nvidia says Super VL is among the best open models on the OCRBench v2 leaderboard.
For companies, an open model that can be fine-tuned on internal data and run on their own servers offers a way to tailor agents to specific processes. Nvidia points to possible uses such as screening insurance claims for fraud or assessing creditworthiness according to a financial firm’s own criteria. It provides the NeMo framework, Python pipelines and open datasets for customization.
Local versions come with a hardware tradeoff
Nemotron 3.5 Bolt 74B-A9B 24GB is a smaller agent model built for local machines. It has 74 billion parameters, with 9 billion active per step, and uses 2-bit weight quantization to fit within 24 GB of memory. Nvidia describes uses such as multi-step tasks, tool calls, web searches, coding and checking whether PC components are compatible.
The model is optimized for Nvidia systems and also supports llama.cpp, Ollama, LM Studio and vLLM. But the 24 GB memory requirement places it within reach mainly of developers and people with powerful workstations, according to the source article.
A smaller Bolt E4B 4GB model is planned for later this month. It is intended for voice assistants, interactive characters and device interfaces, with support for multi-turn conversations, tool use and awareness of the current app. It has not shipped yet.
Nvidia says organizations are already adapting Super VL for areas including finance, document processing and cybersecurity. Examples in the source include receipt processing, acquisition due diligence, credit agreement analysis and threat detection. The company is releasing model weights, datasets and training recipes, while optimizing its models and software for Nvidia hardware.