Nemotron 3.5 Bolt: NVIDIA Introduces 74B Model Targeting 24GB of Memory
NVIDIA has released the multimodal Nemotron 3.5 Super VL and local Bolt 74B-A9B model, which targets 24GB of memory through 2-bit quantization. The smaller Bolt E4B is scheduled to follow on October 23.
Nemotron 3.5 Bolt expands NVIDIA’s model lineup for local agentic tasks. On October 15, the company introduced the Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB models. The smaller Nemotron 3.5 Bolt E4B 4GB variant is not yet available; NVIDIA plans to release it on October 23, 2026.
The company describes the new models as “open models.” In this case, that means publicly available model weights and related artifacts. However, it does not necessarily mean a fully open-source release including all training data and source code.
Nemotron 3.5 Bolt targets 24GB of memory
Nemotron 3.5 Bolt 74B-A9B has 74 billion parameters in total, while activating 9 billion parameters when processing a single token. NVIDIA uses 2-bit quantization to make the model fit within 24GB of memory.
The model is intended for local agentic workflows on NVIDIA RTX systems and DGX Spark. Agentic tasks may include tool use or processing reference materials, while local deployment is particularly relevant where developers and companies are addressing latency, costs, or sensitive data on their own infrastructure.
The 24GB figure describes the model’s memory target, not a universal guarantee of trouble-free operation on every computer with that capacity. Practical use will also depend on inference software support, available system memory, and the specific task.
Multimodal Super VL with a MoE architecture
The concurrently available Nemotron 3.5 Super VL is a Mixture-of-Experts (MoE) multimodal model. It has 120 billion parameters in total, but activates 12 billion of them for each token. Its multimodal focus means it also targets tasks involving documents, images, and video.
NVIDIA presents the model for specialized agents with multimodal reasoning. Developers can therefore adapt the model weights to their own data and deploy them in a controlled environment. The specific benefit, however, will depend on the use case, data, hardware, and integration solution.
Performance claims require independent comparison
In its materials, NVIDIA makes claims about the accuracy, speed, and throughput of the individual models. However, such results are based on benchmarks, specific hardware, and configurations selected by the manufacturer. An independent comparison of the models’ quality, speed, and memory requirements has not yet been confirmed in the available materials.
The company likewise cites expected operational and cost improvements from partner deployments. These figures cannot automatically be applied to all users or different types of workloads.
What to watch next
The next date is October 23, when NVIDIA plans to make Nemotron 3.5 Bolt E4B 4GB available. It will also be important to see what licensing terms apply to the individual models and how quickly compatibility with commonly used local inference tools is added.
For practical evaluation, independent testing of the models outside NVIDIA’s materials will be important, particularly regarding response quality, processing speed, and actual memory usage in local configurations.
Sources
- NVIDIA Blog – Confirms the announcement, availability of Super VL and Bolt 74B-A9B, and the planned release date for Bolt E4B 4GB.
- NVIDIA Nemotron Research – Documents the Bolt 74B-A9B architecture, 24GB memory target, 2-bit quantization, and local focus.
- NVIDIA Technical Blog – Provides the Super VL parameters, multimodal capabilities, and methodology for NVIDIA’s performance claims.
- NVIDIA Glossary – Explains that the open model designation does not necessarily mean complete open-source publication of all data and code.
Verified and updated: October 16, 2026 06:24
