Nvidia Corp. has significantly expanded its artificial intelligence portfolio with the introduction of three new additions to its Nemotron family of open-weights models. Headlining the release is the Nemotron 3.5 Super VL, a massive 120-billion-parameter vision-language model explicitly engineered for enterprise-grade autonomous agents and complex multimodal workflows. Alongside this flagship enterprise offering, the company unveiled two compact iterations dubbed the Bolt models—specifically, the Nemotron 3.5 Bolt 74B-A9B 24GB and the impending Nemotron 3.5 Bolt E4B 4GB. Designed for localized deployment on personal computers and edge infrastructure, these smaller offerings aim to balance low computational overhead with the intelligence required for advanced coding, multistep reasoning, and local research tasks.
The strategic release underscores Nvidia’s continuous push to democratize high-efficiency AI while addressing a rapidly bifurcated market. On one hand, enterprises demand massive, highly capable multimodal engines that can interpret sprawling documents, analyze video feeds, and execute multi-application business workflows securely. On the other hand, developers, researchers, and power users increasingly require localized models that can perform everyday agentic operations directly on consumer hardware without leaking sensitive proprietary data to external cloud servers. By packaging advanced architectures into both high-end server models and ultra-compact local packages, Nvidia is positioning its ecosystem to capture workflows across the entire computing spectrum, from cloud-scale data centers to local laptops and embedded edge devices.
Evolution of the Nemotron Architecture and Background Context
The genesis of the Nemotron family dates back to Nvidia’s broader strategy to establish a robust open-weights portfolio that can compete directly with proprietary models from leading AI labs, as well as prominent open-source ecosystems curated by companies like Alibaba, Meta, and various academic institutions. Earlier this year, in March, the company introduced the Nemotron 3 Super, a 120-billion-parameter text-only model that garnered attention for its high-performance hybrid architecture and expansive context windows of up to one million tokens.
The new Nemotron 3.5 Super VL directly builds upon this foundation, preserving the core structural philosophy of its predecessor while incorporating advanced vision encoders. While the original Nemotron 3 Super was strictly text-bound, the Super VL model integrates comprehensive image and video input processing capabilities—allowing systems to interpret visual data natively within its massive context window.
Underpinning these models is a distinct hybrid architectural design that deviates from traditional transformer-only paradigms. The Nemotron 3.5 family leverages a mixture-of-experts (MoE) framework combined with Mamba-2 state space layers and standard attention mechanisms. This hybrid approach is designed to mitigate the quadratic scaling bottlenecks typically associated with long-context transformer models, enabling faster inference speeds and reduced memory consumption during extended operations. Furthermore, Nvidia utilized sophisticated distillation techniques during the training phase, leveraging outputs from its larger Nemotron 3 Ultra models to supervise and accelerate the learning process of the newer, more efficient variants before executing specialized fine-tuning for software engineering, vision tasks, and autonomous agent behaviors.
Technical Deep Dive: Nemotron 3.5 Super VL Capabilities
The marquee enterprise model, Nemotron 3.5 Super VL, possesses a total parameter count of 120 billion, though its mixture-of-experts routing activates only 12 billion parameters concurrently during any given inference step. This sparse activation mechanism is critical for achieving high throughput without demanding prohibitive computational resources.
According to Kari Briski, Nvidia’s vice president for open models and software, the Super VL model is structured to deliver industry-leading accuracy across multimodal agentic tasks. This includes strict instruction following, complex tool utilization, multi-step business process orchestration, and advanced visual comprehension. Notably, Nvidia’s internal metrics indicate that Super VL achieves performance parity with the company’s largest reasoning models while utilizing roughly one-fifth of the total active parameter count.
To overcome the latency hurdles traditionally inherent in processing extensive language tokens, Nvidia incorporated an innovative open speculative decoding technique known as LiLiCorr. Through speculative decoding, a smaller, highly efficient helper model drafts multiple predicted tokens ahead of the primary execution cycle. The Super VL model then evaluates these drafted tokens in parallel rather than generating each sequential token individually. Nvidia reports that this speculative decoding framework accelerates token generation speeds on DGX Spark systems by up to 2.3 times compared to standard unassisted decoding methodologies.
In practical enterprise applications, the model’s visual understanding extends deeply into document workflows, optical character recognition (OCR), table extraction, and complex chart interpretation. On CharXiv, a widely recognized benchmark for chart reasoning, Super VL achieved an 80.8% accuracy score—substantially outperforming competing models like Nano Omni, which logged 63%. When supplemented with Python code interpretation capabilities, Super VL’s performance on visual reasoning tasks scales further to 84%.
Competitive Benchmarking and Performance Analysis
In evaluations published by Nvidia, the Nemotron 3.5 Super VL demonstrates a nuanced competitive profile when measured against peer open models of comparable scale, such as Alibaba’s Qwen series and models developed by independent research labs.
When pitted against Alibaba’s Qwen3.5-122B-A10B—a model boasting 122 billion total parameters with 10 billion active—Super VL exhibits mixed performance depending on the specific evaluation vector. In rigorous scientific reasoning and terminal-based command execution tasks, Nvidia’s model holds a definitive edge. For instance, on Terminal-Bench 2.1, Super VL scored 66.7, outperforming Qwen3.5’s score of 48.7. Conversely, Qwen3.5 maintains a slight advantage on software-engineering-specific benchmarks such as SWE-bench Verified, registering 71.9 compared to Super VL’s 71.1.
When compared against newer, larger alternatives such as the Qwen3.8-Flash-Next—which incorporates auxiliary embedding and prediction modules that elevate its total footprint to approximately 180 billion parameters—Super VL trails on certain raw benchmark metrics. However, Nvidia emphasizes that Super VL was optimized primarily for execution speed and operational efficiency rather than raw benchmark saturation. Internal company data highlights that Super VL completes 10,000 tasks on SWE-bench Pro up to four times faster when deployed across a cluster of four B200 GPUs, showcasing its viability for high-throughput enterprise production pipelines where latency directly translates to operational expenditure.
Furthermore, on behavioral evaluations managed through Hermes Bench and LangChain frameworks, Super VL secured notable advantages, scoring 63.44% versus 56.33% for competing architectures, and 87.05% versus 84.3% on LangChain metrics, respectively.
The Bolt Models: Localized Efficiency and Edge Computing
While the Super VL targets centralized data centers and heavy enterprise infrastructure, the simultaneous rollout of the Bolt model series addresses the surging demand for powerful, localized artificial intelligence.
The primary release in this category, the Nemotron 3.5 Bolt 74B-A9B 24GB, is directly derived from the Super VL lineage. It utilizes a hybrid mixture-of-experts architecture that activates 9 billion of its total 74 billion parameters during inference. To make such a large model viable for local execution on consumer and workstation hardware, Nvidia utilized aggressive two-bit quantization techniques. This advanced compression successfully reduces the model’s memory footprint to 24 gigabytes, allowing it to fit comfortably within the memory limits of high-end consumer GPUs and dedicated edge hardware like RTX Spark and DGX Spark systems.
In benchmarking evaluations, the Bolt 74B model punches significantly above its weight class. On Terminal-Bench 2.1, the model scored 51.0, outperforming competing architectures such as Qwen3.6-35B-A3B (which scored 44.4) and Gemma4 26B-A4B (scoring 37.2) when tested under baseline BF16 conditions. Nvidia has confirmed that the model is being optimized to integrate seamlessly with popular local execution runtimes, including llama.cpp, Ollama, LM Studio, and vLLM.
Complementing the 74B variant is the smaller Nemotron 3.5 Bolt E4B 4GB, scheduled for release later in the month. Tailored specifically for resource-constrained environments—such as GeForce RTX-equipped consumer PCs, Jetson modules, and embedded edge systems—this ultra-compact model is designed to power real-time conversational assistants, interactive non-player characters (NPCs) in gaming environments, and desktop application agents that require continuous multi-turn tool interaction and state awareness.
Industry Implications and Commercial Availability
The release of the Nemotron 3.5 family reflects a broader maturation phase in the generative artificial intelligence sector. Rather than focusing solely on increasing raw parameter counts, leading developers are increasingly prioritizing architectural innovations—such as hybrid state-space models, advanced quantization, and speculative decoding—to make powerful AI economically and computationally sustainable.
By offering these models under the OpenMDW 1.1 license, which permits commercial use, modification, and redistribution, Nvidia is actively encouraging enterprise adoption and community-driven refinement. The strategy enables commercial entities to deploy sophisticated multimodal reasoning agents on-premises or within private cloud environments, ensuring data privacy and compliance with stringent regulatory frameworks.
All three models—Nemotron 3.5 Super VL and the 74B Bolt variant—are available for immediate download via Hugging Face, with the lightweight Bolt E4B 4GB model set to roll out to the developer community in the coming weeks. As organizations increasingly transition from experimental generative AI pilots to fully autonomous agentic workflows, Nvidia’s latest offerings provide a versatile, high-performance toolkit designed to meet the rigorous demands of modern computing environments.
