NVIDIA has officially launched Nemotron 3.5 Super VL, marking a significant advancement in open vision-language models for enterprise applications. This sophisticated model is engineered to serve as the backbone for continuous, intelligent agents, capable of handling a diverse range of data inputs and tasks.
A Glimpse into its Architecture and Abilities
At its core, Nemotron 3.5 Super VL is a 120-billion-parameter mixture-of-experts (MoE) model. A key feature of its design is the dynamic activation of 12 billion parameters per processing token, optimizing efficiency without sacrificing power. This singular model is equipped to interpret text, analyze images, process documents, and even understand video content. Beyond comprehension, it excels at complex reasoning, interacting with tools, and generating code, providing a comprehensive solution that traditionally required separate specialized models for developers.
Alongside this flagship model, NVIDIA also introduced two more compact, edge-optimized versions: Nemotron 3.5 Bolt 74B-A9B and Bolt E4B, catering to different deployment needs.
Unlocking Performance and Efficiency
The Nemotron 3.5 Super VL boasts an impressive context window, supporting up to 1 million tokens. It operates efficiently on a range of NVIDIA hardware, including a single DGX Spark, one H200, two H100 GPUs, or one GB200, when utilizing the optimized NVFP4 checkpoint. A BF16 checkpoint is also available for broader compatibility.
NVIDIA claims the model delivers strong performance in critical areas like advanced reasoning, automated coding, and document analysis. Notably, the company reports up to five times greater single-GPU throughput compared to other open-source alternatives. While demonstrating exceptional performance on benchmarks such as GPQA-Diamond with a score of 89.90, the model also showed a lower score of 17.78 on the tau3-bench banking task, indicating areas for potential refinement in specific, long-horizon agentic scenarios.
Commitment to Open Source Development
Nemotron 3.5 Super VL is released as an open multimodal agent model, with its weights provided under the OpenMDW-1.1 license. This open approach extends to supplementary resources, as NVIDIA has also made available open datasets, comprehensive training recipes, and reinforcement learning environments within its NeMo Gym framework, fostering community innovation.
The model accepts various inputs, including standard text, JPEG and PNG images, and MP4 video clips up to two minutes in duration. It generates text outputs, with the flexibility to toggle reasoning capabilities through its chat template.
Innovative Architectural Design for Speed
The internal architecture of Nemotron 3.5 Super VL features a hybrid backbone that ingeniously interleaves Mamba-2, mixture-of-experts, and traditional attention layers. Each MoE layer incorporates 512 routed experts. Visual data, including images, document pages, and video frames, are channeled into the language backbone via a C-RADIO vision encoder and a specialized connector.
The model’s speed is significantly enhanced by several cutting-edge techniques:
- Multi-token prediction: A native drafting mechanism proposes multiple tokens, which are then verified in a single forward pass.
- LiLiCorr: A speculative drafter specifically for DGX Spark, offering up to 2.3 times faster token generation.
- NVFP4 with quantization-aware distillation: Utilizes 4-bit weights and activations on Blackwell architectures, and 4-bit weights with 16-bit activations on other platforms.
- Context reuse: Employs prefix caching to maintain both attention caches and Mamba states across agent interactions.
Nemotron 3.5 Super VL was developed through a distillation process from Nemotron 3 Ultra, having been trained on an estimated 26 trillion tokens. Post-training involved approximately 3.8 million samples and reinforcement learning across numerous environments, with an additional 10 billion Indic tokens added during continued pretraining.
Impressive Benchmark Results
Based on NVIDIA-reported scores from the NVFP4 checkpoint, Nemotron 3.5 Super VL demonstrates strong competitive performance:
- Reasoning (GPQA-Diamond): 89.90
- Coding (SWE-bench Verified): 71.07
- Web agent (BrowseComp): 77.5
- Instruction following (IFBench): 71.50
- Charts (CharXiv reasoning): 79.88
- Documents (MMLongBench-Doc): 62.82
In agent harness evaluations, Super VL outperformed its predecessor, Nemotron 3 Ultra, despite Ultra possessing significantly more parameters (approximately 4.6 times more). It achieved 63.44% on Hermes Bench compared to Ultra's 56.33%, and 87.05% on LangChain evaluations versus Ultra's 84.3%. Furthermore, NVIDIA reports that the model can generate over 600 output tokens per second on agentic tasks, showcasing a five-fold speed advantage over models like Inkling Small in certain scenarios.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
