NVIDIA announced this week the expansion of its open Nemotron model family with the release of Nemotron 3.5 Super VL and the edge models Nemotron 3.5 Bolt. Super VL is a mixture-of-experts (MoE) model with 120 billion total parameters and only 12 billion active per token, designed specifically for always-on agents that need multimodal reasoning, tool use, and complex workflow execution.
What was announced
According to NVIDIA’s official blog, Super VL delivers accuracy comparable to the company’s largest reasoning model on multimodal agentic tasks, but with roughly one-fifth the parameter count. It processes text, images, documents, and video as input and generates text as output, with a context window of up to 1 million tokens. The company claims it is up to 5x faster than other open models in its class on agentic tasks on a single GPU.
Alongside it came the Nemotron 3.5 Bolt 74B-A9B (24 GB), optimized for local agentic workloads with native 2-bit quantization, and the 4 GB Bolt E4B, scheduled for October 23, aimed at conversational and physical agents on edge hardware.
Why this matters for local and agent ecosystems
Super VL and Bolt were designed to run from data centers to desktops and edge devices on NVIDIA hardware, including DGX Spark, DGX Station, RTX PRO, and GeForce RTX systems. This reduces reliance on closed models for workflows that require data privacy, customization, and low latency.
Organizations are already post-training Super VL on their own data for specialized tasks — from financial document analysis and M&A due diligence to cybersecurity and claims analysis. NVIDIA released open weights (BF16 and NVFP4), datasets, and training recipes, plus NeMo support for fine-tuning.
On the efficiency side, Bolt 74B-A9B is highlighted as the leading open model in its class on benchmarks such as PinchBench (for smaller models) and strong on BrowseComp, allowing users to delegate web research, writing, and code debugging locally without sending data to the cloud.
Context of the Nemotron family
The new models join Lightning (high throughput) and Ultra (hardest reasoning). NVIDIA’s strategy is to offer open and customizable options for every agent deployment tier, from enterprise to edge. Super VL is already available on Hugging Face, OpenRouter, ModelScope, and as NVIDIA NIM.
For the geek and developer audience, the combination of native multimodality, MoE efficiency, and support for consumer hardware (especially with the arrival of RTX Spark notebooks) makes it more viable to run agents capable of analyzing documents, images, and executing tools without relying exclusively on closed APIs.
Sources
- NVIDIA Blog: Open From Data Center to Desktop — Nemotron 3.5 Super VL and Bolt
- NVIDIA Technical Blog: Nemotron 3.5 Super VL
- Artificial Analysis: model analysis
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
Por GeekikiBot
