NVIDIA Nemotron 3.5 Super VL: open 120B agent model, 12B active

NVIDIA released Nemotron 3.5 Super VL on 15 October 2026, an open multimodal model for agents that read documents, call tools and run for long periods. The model has 120 billion parameters, of which 12 billion are active for each token, and takes text, images, documents and video. Two smaller Bolt models for local machines came with it. Everything below is NVIDIA's account unless it says otherwise.

What was released

• Nemotron 3.5 Super VL (120B-A12B): available now on Hugging Face, OpenRouter, the ModelScope community and as an NVIDIA NIM microservice, according to NVIDIA's launch blog.
• Nemotron 3.5 Bolt 74B-A9B 24GB: available now on Hugging Face. NVIDIA says it uses native 2-bit quantisation to fit in 24GB of memory and works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
• Nemotron 3.5 Bolt E4B 4GB: a small model for voice assistants and embedded devices, due on 23 October 2026.

The Hugging Face model card lists the licence as OpenMDW 1.1 and a context length of up to 1 million tokens. It says the BF16 example runs on 4 H100 GPUs, and points to a separate NVFP4 version for DGX Spark and data-centre GPUs. NVIDIA says it also released training datasets and recipes.

What NVIDIA claims about speed and accuracy

NVIDIA says the model matches its much larger Nemotron 3 Ultra on agent tasks at about one-fifth of the parameters. In its technical blog, Super VL scores 63.44% on Hermes Bench against 56.33% for Ultra, and 87.05% against 84.3% on LangChain behavioural evaluations. It scored 80.8% on the CharXiv chart-reasoning benchmark, or 84% with a Python interpreter.

On speed, NVIDIA says a single GPU gives up to 5 times the throughput of other open models of similar size and that tasks finish 4 times faster. The blog also says an independent test by EXO Labs on a DGX Spark placed the model on the accuracy-versus-completion-time frontier. The Decoder, reporting on the launch, notes that on the coding benchmark Super VL is slightly less accurate than competitors and that the model trails the leading closed models on Artificial Analysis's overall ranking.

Customer forecasts and reported accuracy

NVIDIA names several organisations that post-trained the model on their own data. Scale AI built a vendor-credit-processing agent for Skydance, for which NVIDIA says inference cost will decrease by 40x and latency by 5x. NVIDIA reports that Ramp expects up to 50% higher throughput and 10x lower latency converting receipt images to structured data.

NVIDIA also reports that AI21 post-trained the model for Alice’s threat-intelligence and vulnerability analysis, improving accuracy by 40%.

The wording matters: “will decrease” and “expects” make the Skydance and Ramp efficiency figures forecasts, whereas NVIDIA describes Alice’s accuracy improvement as a post-training result. The baselines are not stated in the accompanying text. All of these customer figures are vendor-reported and concern customised models, not independent tests of the released weights. The blog names EY, Hebbia, Atlassian, CrowdStrike and Salesforce as customising or integrating the model, but gives no results for them.

What a team should check

• Licence: read OpenMDW 1.1 against your use before building on it; the model card calls the licence 'other', not Apache or MIT.
• Hardware: 4 H100 GPUs is the card's example for BF16 at 1M tokens; NVIDIA notes deployments may use a shorter context to fit memory.
• Your own task set: the Hermes and LangChain numbers come from agent harness evaluations NVIDIA ran or cites, not from your documents.
• Unknown from the sources: price for the NIM service, and independent results for the Bolt models.