IT之家 October 15 news: NVIDIA has announced the launch of the open multimodal model Nemotron 3.5 Super VL, aimed at AI agents that need to run continuously and perform multi-step tasks.
The model uses a Mixture of Experts (MoE) architecture with a total of 120 billion parameters, of which 12 billion are activated. According to NVIDIA, when performing agentic tasks on a single GPU, the model's throughput can reach up to 5 times that of other open models of similar size.
According to the introduction, Nemotron 3.5 Super VL integrates reasoning, instruction following, and tool calling capabilities with text, image, document, and video understanding capabilities in a single model, and can be used for scenarios such as code testing and debugging, cross-website information retrieval, document analysis, cybersecurity threat identification, and video content retrieval. Developers can use the NVIDIA NeMo toolchain to perform post-training with their own data, tools, policies, and workflows, and adjust model behavior according to business needs.
NVIDIA also announced multiple enterprise customization cases. For example, AI data technology provider Scale AI customized a vendor credit processing agent for the media entertainment company Skydance Corporation, with inference costs for related workloads expected to drop by 97.5% and latency reduced by 80%.
The 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 reduced by 90%. AI21 partnered with AI security company Alice to customize a model for threat intelligence and vulnerability analysis, improving accuracy on related tasks by 40%.
In terms of model evaluation, NVIDIA said Nemotron 3.5 Super VL scored 80.8% on the CharXiv chart reasoning benchmark, higher than Nemotron 3 Nano Omni's 63%; when combined with Python code interpretation, the score reached 84%.
On the Hermes Bench and LangChain behavioral evaluations, the model achieved 63.44% and 87.05% respectively, higher than Nemotron 3 Ultra's 56.33% and 84.3%. NVIDIA also stated that the model is among the top-performing open models on the OCRBench v2 leaderboard.
To reduce inference overhead, Nemotron 3.5 Super VL adopts techniques such as speculative decoding, mixed precision, and context reuse. According to NVIDIA, 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 the NVIDIA Blackwell architecture; on non-Blackwell platforms, the same checkpoint uses 4-bit weights and 16-bit activations.
NVIDIA also announced deployment optimization solutions for distributed inference. According to its disclosed data, under test conditions using 16 NVIDIA B200 GPUs with 50 tokens generated per second per user, NVIDIA Dynamo combined with cache-aware routing, session affinity, and KV cache offloading can serve up to 8 times the number of users and reduce first-token latency by 85%.
This release also includes base model, post-trained model, and NVFP4 quantized model weights, as well as training recipes, datasets, and reinforcement learning environments.
NVIDIA stated that Nemotron 3.5 Super VL used approximately 26 trillion tokens in total during pre-training and continued pre-training phases, approximately 3.8 million samples during post-training, and conducted reinforcement learning in dozens of environments. The newly released datasets cover areas such as agentic tool calling, financial reasoning, instruction following, long context, fairness, and vision-language understanding.
In addition to Super VL, NVIDIA also launched two Nemotron 3.5 Bolt on-device models. Nemotron 3.5 Bolt 74B-A9B 24GB uses an MoE architecture with 74 billion total parameters and 9 billion activated parameters, compressing the model to 24GB through native 2-bit quantization. It is aimed at local agentic tasks and supports llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. Nemotron 3.5 Bolt E4B 4GB, targeted at scenarios such as local voice assistants, interactive characters, and embedded interfaces, is planned for release on October 23.
IT之家 notes that Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB are now openly available; developers can access the relevant model resources through channels such as Hugging Face, OpenRouter, and ModelScope. Super VL can be deployed on DGX Spark, DGX Station, RTX systems, data centers, or the cloud; Bolt E4B 4GB is planned for release on October 23.
