NVIDIA released two new AI models on October 15, 2026. They are called Nemotron 3.5 Super VL and Nemotron 3.5 Bolt.
Both models are open-weight. Developers can download and run them directly. That is confirmed by NVIDIA’s developer blog.

Nemotron 3.5 Super VL brings multimodal reasoning to agents
It has 120 billion total parameters. Only 12 billion are active per token.
It runs on a hybrid Mamba-Transformer backbone. A C-RADIO vision encoder gives it image and video understanding, per NVIDIA.
NVIDIA distilled Super VL from its larger Nemotron 3 Ultra model. Training used roughly 26 trillion tokens, plus post-training on about 3.8 million samples.
The model also went through reinforcement learning. NVIDIA ran this across dozens of simulated environments.
Super VL supports a 1-million-token context window. It reads text, images, and documents natively. It also handles up to two minutes of video.
That removes a separate vision-language model from most agent pipelines. One model now handles what used to require two, according to NVIDIA.

Benchmark scores show a steady climb against its predecessor
NVIDIA’s own benchmark numbers tell a layered story. On SWE-bench Verified, Super VL scores between 71.07 and 71.10.
On GPQA-Diamond, it reaches 89.90. On Terminal-Bench 2.1, it scores 60.80.
On BrowseComp, the score climbs to 77.5. On OSWorld-Verified, it lands around 51.9 to 52.3, per NVIDIA’s reported figures.
The comparison against Nemotron 3 Ultra matters most. On Hermes Bench, Super VL scores 63.44% versus Ultra’s 56.33%.
On LangChain evals, Super VL hits 87.05% against Ultra’s 84.3%. It does this with about one-fifth the active parameters, NVIDIA says.
That is not a small gain. That is a smaller model beating its own predecessor.
Coding speed and the smaller Bolt companion model
NVIDIA ran SWE-bench Pro across 10,000 tasks. Using four B200 GPUs, Super VL finished in roughly 10 hours.
Comparison models of similar accuracy took far longer. NVIDIA says 50 to 120 hours. That is roughly a 4x speed advantage.
Alongside Super VL, NVIDIA released Nemotron 3.5 Bolt the same day. Bolt is smaller, with 74 billion total parameters and 9 billion active.
Bolt is built to run agentic workloads locally. A single 24GB VRAM workstation or laptop GPU is enough, NVIDIA says.

Why this open-weight release matters for AI agents
Weights are downloadable on Hugging Face. Training data and code stay closed, per The Decoder.
That makes this a partial-openness release. Commercial use may face some licensing restrictions.
Independent evaluator Artificial Analysis reportedly ranks Super VL at 26. That score is on its Intelligence Index. It reportedly ties it for the top US open-weight model at release.
NVIDIA has not published pricing. No hosted-API cost is confirmed yet.
This release joins a wave of recent open-weight launches. These include Mistral AI’s Mistral Large 4 and Alibaba’s Qwen3.8-Omni-Flash. Enterprise interest in AI agent deployment keeps growing too.
Developers should watch Hugging Face download counts over the coming weeks. Independent benchmark trackers will also reveal how Super VL and Bolt perform outside NVIDIA’s own test suite.
