NVIDIA Releases Open 120B Nemotron 3.5 Super VL Model

What Is Nemotron 3.5 Super VL?

NVIDIA has released Nemotron 3.5 Super VL, an open-source vision-language model built for always-on enterprise agents¹. The model has 120 billion parameters in total, but only 12 billion are active for each token it processes. This is possible because it uses a mixture of experts (MoE) design, where many smaller sub-models share the work and only the relevant ones activate for any given input. The system can read text, images, documents, and short video clips, then reason, call external tools, and write code—all in one model. For developers, this means one open checkpoint replaces the need for a separate text model and a separate vision model.

How Does It Work?

The model’s backbone interleaves Mamba-2 layers, MoE layers, and traditional attention layers. Each MoE layer contains 512 routed experts, meaning the system can pick the best specialists for a task on the fly¹. A custom vision encoder called C-RADIO processes images, document pages, and video frames, then passes them to the language backbone.

Speed comes from four techniques. Multi-token prediction lets a native drafter propose several words at once, which are verified in a single forward pass. LiLiCorr is a speculative drafter optimized for Blackwell GPUs like the DGX Spark, boosting token generation up to 2.3× faster. NVFP4 is a quantization-aware distillation method that compresses weights and activations to 4-bit precision on Blackwell hardware, while using 4-bit weights with 16-bit activations on other chips. Finally, context reuse (prefix caching) keeps both the attention cache and the Mamba states across different turns of an agent conversation, saving computation¹.

Nemotron 3.5 Super VL was distilled from the larger Nemotron 3 Ultra model. The training data covered roughly 26 trillion tokens, followed by about 3.8 million post-training samples and reinforcement learning across dozens of environments¹. An extra 10 billion Indic-language tokens were added during continued pretraining to improve multilingual capability.

Performance and Deployment

On reasoning benchmarks, the NVFP4 checkpoint scored 89.90 on GPQA-Diamond, a graduate-level science question set, and 71.07 on SWE-bench Verified, which measures real-world coding ability¹. In agent-focused evaluations, Super VL outperformed the older Nemotron 3 Ultra despite having far fewer total parameters: 63.44% on Hermes Bench versus Ultra’s 56.33%, and 87.05% on LangChain evals versus 84.3%

The model is available on Hugging Face, OpenRouter, and ModelScope, and can be run through NVIDIA NIM. Supported runtimes include vLLM, SGLang, and TensorRT-LLM on Blackwell and Hopper GPUs. Local paths such as llama.cpp and Unsloth also work. For production use, Dynamo recipes add key-value-aware routing, which NVIDIA says can serve eight times more users at 50 tokens per second per user while cutting first-token latency by 85%

Early customer results include Scale AI reporting 40× lower inference cost and 5× lower latency for vendor credit processing; Ramp achieving up to 50% higher throughput and expecting 10× lower latency for receipt-to-JSON extraction; and AI21 seeing a 40% accuracy boost on an Alice threat-intelligence task¹.

Alongside Super VL, NVIDIA released two smaller edge models: Nemotron 3.5 Bolt 74B-A9B, which targets local agents and fits into 24 GB of memory thanks to native 2-bit quantization, and the tiny Bolt E4B 4GB version, aimed at voice assistants and embedded agents on GeForce RTX, Jetson, and DGX Spark systems.