NVIDIA Releases a 120-Billion-Parameter AI Model and Two Edge Computing Models

en.Wedoany.com Reported - On October 15, U.S.-based NVIDIA announced the expansion of its Nemotron 3.5 open model series, introducing the 120-billion-parameter Nemotron 3.5 Super VL multimodal model, along with two models for local computing devices: Nemotron 3.5 Bolt 74B-A9B and Bolt E4B. The Super VL and 74B versions are now officially available, while the 4GB version is scheduled for release on October 23. The three models target enterprise-grade AI agents, personal computing devices, and embedded edge terminals, respectively.

Super VL was developed by NVIDIA in collaboration with the Nemotron Coalition, supporting instruction execution, tool calling, multi-step business processing, and visual understanding. NVIDIA's published test results show that the model's multimodal task accuracy can reach a level comparable to its large reasoning models, with a parameter scale about one-fifth of the latter. In performance comparisons among similar open models, its inference throughput reaches up to 5 times that of competing models. The model can process web pages and PDF files containing text, images, charts, and tables, and supports customized training based on enterprise internal data.

Super VL was trained using approximately 26 trillion pre-training and continued pre-training tokens, along with approximately 3.8 million post-training samples, optimized through multiple reinforcement learning environments. NVIDIA has simultaneously open-sourced the model weights, training methods, and datasets covering tool calling, financial operations, long-context processing, safety, and visual language understanding. Developers can fine-tune and conduct reinforcement learning training via NVIDIA NeMo and NeMo Gym, and deploy models using NVIDIA NIM microservices.

The Bolt 74B-A9B, designed for local operation, adopts a mixture-of-experts architecture with a total parameter count of 74 billion and approximately 9 billion parameters activated per inference. Through native 2-bit quantization technology, it keeps the model memory required for operation at 24GB, and can be deployed on GeForce RTX computers, DGX Spark, DGX Station, and RTX PRO workstations. NVIDIA has provided runtime support for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM, allowing users to perform agent tasks such as code writing, software debugging, and information retrieval locally.

The other model, Bolt E4B, targets 4GB memory configurations and is compatible with GeForce RTX, Jetson, and RTX Spark platforms, supporting multi-turn conversations, voice assistants, interactive characters, and embedded intelligent terminals. NVIDIA positions it for physical AI and conversational agent applications, covering consumer devices, in-vehicle interaction systems, and other terminals requiring local inference.

NVIDIA also announced progress on enterprise customization and software integration for Super VL. Scale AI has completed post-training for Skydance's vendor credit financing agent and EY's M&A due diligence agent; AfterQuery developed a credit agreement analysis model for Hebbia; and Mercor used APEX-SWE code evaluation data for model training. Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce, and Synopsys have also respectively carried out integrations related to search, security analysis, engineering design, government processes, and semiconductor design verification.

In inference deployment, Super VL supports the NVIDIA Dynamo runtime environment. NVIDIA's published performance data indicates that this combination can support up to 8 times the number of users under the corresponding test configuration, while reducing latency by 85%. The model is now available through Hugging Face, OpenRouter, ModelScope, and NVIDIA NIM. Enterprises can deploy it on their own servers, data centers, or cloud computing environments, and continue training and adjusting the model according to internal business processes.