Nvidia on Thursday unveiled three new members of its Nemotron family: a 120-billion-parameter vision-language model called Nemotron 3.5 Super VL for enterprise agents, plus two compact Bolt models designed to run coding, research, and conversational tasks locally on laptops. The company positioned all three releases around speed and efficiency, with the Bolt models specifically targeting everyday agentic tasks that require multistep reasoning, tool calling, and web browsing on consumer hardware.
The flagship Super VL combines reasoning and tool use with image and document understanding across 120 billion total parameters, though only 12 billion are active during inference. Nvidia reports that Super VL achieved 66.7 on Terminal-Bench 2.1 compared with Alibaba's Qwen3.5-122B-A10B at 48.7, and scored 63.44% versus 56.33% on Hermes Bench against Ultra. However, it trailed Qwen3.5 on SWE-bench Verified, 71.1 to 71.9, and fell behind Qwen3.8-Flash-Next on all five benchmark tests. On CharXiv chart-reasoning tasks, Super VL delivered 80.8% accuracy versus Nano Omni's 63%, with Python code interpretation raising that figure to 84%. The company also reported fourfold faster completion of 10,000 SWE-bench Pro tasks on four B200 GPUs, though with lower accuracy than Qwen3.8-Flash-Next and DeepSeek V4 Flash.
The larger Bolt model, Nemotron 3.5 Bolt 74B-A9B 24GB, activates 9 billion of its 74 billion parameters at a time and achieved 51.0 on Terminal-Bench 2.1, outperforming Qwen3.6-35B-A3B at 44.4 and Gemma4 26B-A4B at 37.2. Nvidia says the company is optimizing the model for RTX Spark and DGX Spark systems, with support for llama.cpp, Ollama, LM Studio, and vLLM. According to Kari Briski, Nvidia's VP for open models and software, the large model is meant to deliver "leading accuracy on multimodal agentic tasks — instruction following, tool use, multistep business workflows, and visual understanding — comparable with NVIDIA's largest reasoning model at roughly one-fifth the parameter count." A second, even smaller version called Nemotron 3.5 Bolt E4B 4GB is scheduled for later this month, targeting conversational and physical agents on GeForce RTX, Jetson, and Spark hardware.
The report notes Nvidia achieved the 24GB memory footprint for Bolt through two-bit quantization to compress the 74-billion-parameter model, using mixed precision that includes two-bit experts. The Nemotron coalition continued with the 3.5 family's hybrid architecture combining Mamba-2, attention, and mixture-of-experts layers, but added a vision encoder for Super VL. To accelerate text generation, Nvidia deployed LiLiCorr, a technique for speculative decoding where a smaller helper model drafts several tokens ahead that Super VL checks together rather than generating each token individually, speeding up token generation on DGX Spark by up to 2.3 times compared with decoding without speculation. The company also continued training the language model using distillation from Nemotron 3 Ultra before further training for software engineering, vision, and agent tasks.
The new models are now available for download on Hugging Face under the OpenMDW 1.1 license, which permits commercial use and modification, with the two larger models available immediately and the smallest Bolt E4B 4G model launching in the near future. The releases position Nvidia to compete in both the enterprise multimodal agent market and the growing segment of local, on-device AI applications requiring minimal memory overhead. By enabling a 74-billion-parameter model to run in 24GB of memory, Nvidia has made enterprise-grade reasoning accessible to hardware configurations that previously could only handle much smaller models. The strategy reflects a broader industry push toward deployment flexibility, where organizations can choose between cloud-based power and local privacy depending on their specific workflow requirements.