NVIDIA Pushes Agent Models to the Desktop: 120B Super VL Matches Its Largest Reasoning Model, Edge Bolt Fits in 24GB

NVIDIA widened its open agent lineup from a single tier into three covering data center to desktop, so organizations no longer have to choose between capability and deployability. The Nemotron family previously had Lightning, released in August, for high-throughput workloads and Ultra for the hardest reasoning tasks, with no counterpart on the edge side. Super VL and Bolt 74B-A9B are available now and Bolt E4B is set for Oct. 23, together covering large enterprise servers down to edge devices and local hardware.

Super VL compresses multimodal agentic capability into a smaller parameter budget, letting enterprise agents finish high-volume tasks faster across workflows spanning days or weeks. Against NVIDIA's largest reasoning model, Super VL reaches comparable accuracy on instruction following, tool use, multistep business workflows and visual understanding at roughly one-fifth the parameter count. NVIDIA describes Super VL as the fastest open agentic model in its class, 5x faster than the next best.

Bolt 74B-A9B 24GB uses native 2-bit quantization to fit in 24GB of memory and run alongside other applications on RTX/DGX Spark. Among models with a similar memory footprint it balances intelligence and speed: smart enough for most jobs, yet much faster than the smartest models. It is the #1 open model in its class on PinchBench and the #2 open model for BrowseComp.

Super VL ships with open weights, open datasets and training recipes, so organizations can post-train on their own data and deploy on their own infrastructure. Unlike closed frontier models, organizations can post-train, deploy and oversee what they build themselves. Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments; Scale AI, AfterQuery, AI21 and Alice, Mercor and ModelScope are among those customizing it.

This report is for researchers and engineers who need to run agents on their own infrastructure, and it covers the capability positioning, availability dates and deployment tiers of the three open models Super VL, Bolt 74B-A9B 24GB and Bolt E4B 4GB. Super VL and Bolt 74B-A9B are available now and Bolt E4B is set for Oct. 23; Super VL is available on Hugging Face, Open Router, the ModelScope community and as NVIDIA NIM, while Bolt 74B-A9B 24GB is on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

Accuracy, speed multiples and leaderboard ranks are NVIDIA's statements in the release, and the material provides no independent evaluation or item-by-item comparison with external models. Bolt E4B 4GB is not yet released, so only its positioning is described. The accuracy comparison with NVIDIA's largest reasoning model rests on the roughly one-fifth parameter count, and the material lists no specific evaluation sets or scores.