Nemotron 3.5 Super VL: Nvidia's Open Model for AI Agents

Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters. It is designed for enterprise agents that run in the background without pause, working through text, images, documents and video. Two smaller "Bolt" models for local PCs and edge devices launch alongside it.

The release adds to a growing field of open-weight models from large tech companies. Nvidia's version comes with more of the underlying materials than most.

Architecture and claimed performance

Super VL uses a Mixture-of-Experts design. Only 12 billion of its 120 billion parameters are active for each token. Nvidia built the model with the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs.

The model was distilled from Nemotron 3 Ultra, Nvidia's largest reasoning model. Nvidia's main claims:

Matching Ultra at a fraction of the size. On instruction following, tool calls, multi-step business processes and image analysis, Super VL should reach Ultra's accuracy with roughly one-fifth of the parameters. In agent tests from Hermes and LangChain, Nvidia says it scores higher than Ultra.
Better coding. It beats its predecessor, Nemotron 3 Super, on SWE-bench and Terminal-Bench.
Throughput. Citing the Artificial Analysis leaderboard, Nvidia reports up to five times the throughput of similarly sized open models on agent tasks on a single GPU. In an independent EXO Labs test on the DGX Spark, no other model was both faster and more accurate.

Super VL also handles OCR, tables, charts, multi-page documents and video without a separate vision model. It scores 80.8 percent on the CharXiv chart benchmark, compared with 63 percent for the older Nemotron 3 Nano Omni. That rises to 84 percent when the model can run Python code.

How Nvidia cuts the compute bill

Nvidia compresses the model with its 4-bit NVFP4 format. Blackwell GPUs run the model fully in 4-bit precision. Older cards store only the weights at 4 bits and do the math at 16 bits, and sensitive parts of the model keep higher precision. A built-in draft component proposes tokens ahead of time for the main model to verify. On the DGX Spark, Nvidia recommends its LiLiCorr method instead and reports text generation up to 2.3 times faster.

Agents resend the same instructions, tool descriptions and history with every request, so Super VL caches that repeated work. Nvidia's Dynamo inference software routes each session to the GPU that already holds its cache. On 16 B200 GPUs, Nvidia says this allows eight times as many concurrent users and cuts time to first token by 85 percent.

Bolt: local agents, but not cheap ones

Bolt 74B-A9B 24GB is a slimmed-down Super VL with 9 of 74 billion parameters active per step. Aggressive 2-bit quantization fits it into 24 GB of memory. It can write and debug code, search the web, or work out which PC parts suit a system. It supports llama.cpp, Ollama, LM Studio and vLLM. Still, 24 GB means pricey high-end hardware, so the target audience is developers and workstation owners for now. A 4 GB Bolt E4B for voice assistants and device interfaces is due later this month.

Early users and open materials

Nvidia says dozens of organizations are already fine-tuning Super VL, mainly in document-heavy, data-sensitive sectors. Ramp turns receipt photos into structured data. EY uses the model for acquisition due diligence. CrowdStrike is building agents to defend enterprise networks.

Weights ship in base, post-trained and 4-bit versions, together with datasets, training recipes and reinforcement learning environments. One dataset, built with the Indian Institute of Technology Madras and its AI4Bharat lab, holds about 2.83 trillion tokens in 22 Indian languages.

Our Take

For companies that want agents on their own servers, tuned to their own data and outside a provider's terms of service, Super VL looks like a serious option. The open datasets and recipes go further than most weight-only releases. Paired with tooling such as post-training platforms, they lower the barrier to customization.

The motive is clear, though. Every variant is tuned for Nvidia hardware, and more self-hosted models mean more demand for its chips. Meta, Google, Anthropic and OpenAI are building their own chips, so this suggests open models are becoming Nvidia's hedge. It is worth watching whether independent benchmarks confirm Nvidia's numbers, and whether rivals like Reflection Beam can compete on openness.