Blog
NVIDIA releases Nemotron 3.5 Super VL and Bolt open models for agents, from servers to laptops
On 15 October 2026 NVIDIA released Nemotron 3.5 Super VL, a 120B open multimodal model for enterprise agents, and Bolt 74B-A9B, a version compressed to fit in 24GB of memory. A 4GB Bolt model is due on 23 October. Performance claims are NVIDIA's own; Artificial Analysis published independent scores.
NVIDIA released two new open models on 15 October 2026 and announced a third for later this month. Nemotron 3.5 Super VL is a large multimodal model for enterprise agents, and Nemotron 3.5 Bolt 74B-A9B 24GB is a compressed version designed to run on local hardware. A smaller model, Nemotron 3.5 Bolt E4B 4GB, is scheduled for 23 October.
Nemotron 3.5 Super VL
According to NVIDIA's developer blog, Super VL is a mixture-of-experts model with 120 billion parameters, of which 12 billion are active for each token. It combines reasoning, instruction following and tool use with understanding of text, images, documents and video. NVIDIA's pitch is that one model can handle both text and visual tasks, so teams do not need to run separate language and vision models and pass context between them.
NVIDIA says the release includes base and NVFP4-quantised checkpoints along with training recipes, data and reinforcement-learning material. Developers can post-train the model on their own data, tools and policies using NVIDIA NeMo. The model is available on Hugging Face, OpenRouter and ModelScope, and as an NVIDIA NIM.
NVIDIA describes Super VL as delivering "the fastest task completion and highest accuracy" for specialised always-on agents. That is the company's own claim. Independent benchmarking firm Artificial Analysis gave Super VL a score of 26 on its Intelligence Index and ranked it the highest-scoring US open-weight model in its results at the time of release. That ranking predates Meta's Muse Spark Open, which was released the same day.
Nemotron 3.5 Bolt 74B-A9B 24GB
Bolt is derived from the Super VL family and is aimed at running agents on a personal machine. NVIDIA's technical report describes it as a hybrid Mamba-Transformer vision-language model with 74 billion total parameters and 9 billion active. To fit into 24GB of memory, NVIDIA compressed the architecture and quantised the expert weights to about two bits per parameter. Other parts of the model keep higher precision.
NVIDIA says Bolt is tuned for reasoning, tool use, web browsing and terminal work. It is released with a speculative "drafter" model that helps it reach its target speed, which the report puts at 84 tokens per second on NVIDIA's RTX Spark hardware. Bolt is available on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. NVIDIA says it runs on RTX Spark laptops, which arrive on 16 October, as well as GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
Artificial Analysis scored Bolt at 21 on its Intelligence Index, about 1.5 points behind NVIDIA's earlier Nemotron 3 Ultra, which has roughly seven times as many total parameters. The firm also noted that Bolt uses far more output tokens per task than Super VL. That affects speed and cost in real use.
Bolt E4B 4GB
The third model, Nemotron 3.5 Bolt E4B 4GB, is meant for voice assistants, interactive characters and physical agents on devices such as GeForce RTX cards and Jetson. NVIDIA says it is coming on 23 October. It was not available at the time of writing.
Why it matters
The release gives developers a matched set of open models, from a server-class multimodal model down to one that fits on a laptop. Because the weights, recipes and some training data are published, organisations can adapt the models instead of relying only on hosted APIs. NVIDIA's performance claims should still be weighed against independent tests on real workloads.