News · SI
NVIDIA has released Nemotron 3.5 Super VL, a 120B open multimodal model for always-on agents, plus Bolt models that run agents locally in 24 GB and 4 GB. Open weights, datasets and recipes make it a serious option for teams that need to post-train on their own data.
NVIDIA has released Nemotron 3.5 Super VL, an open 120 billion parameter model for always-on agents, alongside two Bolt edge models built to run agents on local hardware[^1].
01 / 04
Super VL handles instruction following, tool use, multi-step business workflows and visual understanding, including OCR, tables, charts, web pages and PDFs[^1]. NVIDIA says it matches its largest reasoning model on these agent tasks at roughly one fifth the parameter count, and is up to five times faster than the next best open model in its class[^1]. Those are vendor numbers, so benchmark it on your own workload before you switch.
It was trained on about 26 trillion tokens and post-trained with roughly 3.8 million samples plus reinforcement learning across dozens of environments[^1].
02 / 04
Bolt 74B-A9B uses native 2-bit quantisation to fit in 24 GB, so it can run research, writing and debugging agents locally on RTX and DGX Spark systems alongside other apps[^1]. Bolt E4B fits in 4 GB and targets voice assistants and embedded interfaces on GeForce RTX, Jetson and Spark hardware[^1]. Super VL and Bolt 74B-A9B are out now; Bolt E4B arrives on 23 October[^1].
03 / 04
NVIDIA is shipping open weights, open datasets and training recipes, plus reinforcement learning environments in NeMo Gym[^1]. Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys are integrating it, and Scale AI, Mercor and others are post-training it for finance, security and coding agents[^1].
Super VL is on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM, and Bolt 74B-A9B runs on llama.cpp, Ollama, LM Studio, Unsloth and vLLM[^1].
04 / 04
The useful part is the tiering. A team can now prototype an agent on a hosted Super VL endpoint, post-train it on internal documents and policies, and push a smaller Bolt model to laptops or devices, all from one open family. If your agents read screenshots, invoices or PDFs, a single multimodal model also removes the hand-off between a vision model and a text model, which is where many agent pipelines lose context.
FAQ
Nemotron 3.5 Super VL is an open 120 billion parameter multimodal model NVIDIA released on 15 October 2026 (US time). It is built for always-on agents and handles instruction following, tool use, multi-step business workflows and visual understanding of web pages, PDFs, tables and charts.
Bolt is NVIDIA's new edge tier. Bolt 74B-A9B uses native 2-bit quantisation to fit in 24 GB and run everyday agent tasks locally, and Bolt E4B fits in 4 GB for fast voice and conversational agents. Bolt 74B-A9B is out now and Bolt E4B is due on 23 October.
Yes. NVIDIA ships open weights, open datasets and training recipes, and supports customisation through NVIDIA NeMo, Python pipelines and reinforcement learning environments in NeMo Gym. Companies including Atlassian, CrowdStrike, Perplexity and Salesforce are already integrating or customising it.
It is available on Hugging Face, OpenRouter, the ModelScope community and as an NVIDIA NIM microservice. Bolt 74B-A9B is on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
NVIDIA says it is the fastest open agentic model in its class, up to five times faster than the next best on throughput, citing the Artificial Analysis leaderboard. It also claims accuracy comparable with its largest reasoning model at about one fifth the parameter count. These are vendor figures, so test on your own workloads.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.
