Article · AI translation
IT Home reported on October 15 that NVIDIA announced the open multimodal model Nemotron 3.5 Super VL, designed for AI agents that need to run continuously and carry out multistep tasks.
The model uses a mixture of experts (MoE) architecture with 120 billion total parameters and 12 billion active parameters. NVIDIA says that, when performing agent tasks on a single GPU, the model can achieve up to 5 times the throughput of other open models of comparable size.
According to NVIDIA, Nemotron 3.5 Super VL combines reasoning, instruction following and tool use with understanding of text, images, documents and video in a single model. It can be used for code testing and debugging, information retrieval across websites, document analysis, cybersecurity threat identification and video content retrieval. Developers can use NVIDIA’s NeMo toolchain to conduct post-training with their own data, tools, policies and workflows, and adjust the model’s behavior to meet business needs.
NVIDIA also announced several enterprise customization examples. For instance, AI data technology provider Scale AI customized a supplier credit processing agent for media and entertainment company Skydance Corporation. Inference costs for the associated workload are expected to fall by 97.5%, and latency by 80%.
Financial operations platform Ramp is converting receipt images into structured JSON data containing merchant, date and payment information, with throughput expected to increase by up to 50% and latency to fall by 90%. AI21 worked with AI security company Alice to customize a model for threat intelligence and vulnerability analysis, improving accuracy on the relevant tasks by 40%.
On model evaluations, NVIDIA says Nemotron 3.5 Super VL scored 80.8% on the CharXiv chart reasoning benchmark, above Nemotron 3 Nano Omni’s 63%. With the addition of a Python code interpreter, its score reached 84%.
The model scored 63.44% on Hermes Bench and 87.05% on the LangChain behavior evaluation, respectively, above Nemotron 3 Ultra’s 56.33% and 84.3%. NVIDIA also says the model ranks among the better-performing open models on the OCRBench v2 leaderboard.
To reduce inference overhead, Nemotron 3.5 Super VL uses techniques including speculative decoding, mixed precision and context reuse. NVIDIA says its LiLiCorr speculative decoding method can increase token generation speed by up to 2.3 times on DGX Spark. The model also supports NVFP4 quantization, using 4-bit weights and 4-bit activations on NVIDIA’s Blackwell architecture; on non-Blackwell platforms, the same checkpoint uses 4-bit weights and 16-bit activations.
NVIDIA also announced deployment optimizations for distributed inference. According to its disclosed data, under test conditions using 16 NVIDIA B200 GPUs and generating 50 tokens per user per second, NVIDIA Dynamo, combined with cache-aware routing, session affinity and KV cache offloading, can serve up to 8 times as many users and reduce time to first token by 85%.
The release also includes base model, post-trained model and NVFP4-quantized model weights, as well as training recipes, datasets and reinforcement learning environments.
NVIDIA says Nemotron 3.5 Super VL used approximately 26 trillion tokens across pretraining and continued pretraining, approximately 3.8 million samples during post-training, and reinforcement learning in dozens of environments. The newly released open datasets cover areas including agent tool use, financial reasoning, instruction following, long context, fairness and vision-language understanding.
Alongside Super VL, NVIDIA introduced 2 Nemotron 3.5 Bolt on-device models. Nemotron 3.5 Bolt 74B-A9B 24GB uses a MoE architecture with 74 billion total parameters and 9 billion active parameters. Native 2-bit quantization compresses the model to 24GB. Designed for local agent tasks, it supports llama.cpp, vLLM, Ollama, Unsloth and LM Studio. Nemotron 3.5 Bolt E4B 4GB is intended for applications such as local voice assistants, interactive characters and embedded interfaces, and is scheduled for release on October 23.
IT Home notes that Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB are already available, and developers can access the model resources through channels including Hugging Face, OpenRouter and ModelScope. Super VL can be deployed on DGX Spark, DGX Station, RTX systems, in data centers or in the cloud; Bolt E4B 4GB is scheduled for release on October 23.