NVIDIA Expands Nemotron 3.5 With 120B Super VL and Compact Bolt AI Models
NVIDIA has expanded its open AI model lineup with Nemotron 3.5 Super VL, a 120 billion parameter model designed for advanced AI agents, alongside 2 smaller models, Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB, optimized for local computing. According to NVIDIA's official announcement, the lineup targets enterprise AI infrastructure, GeForce RTX gaming PCs, and compact edge devices, with deployment memory targets starting at 4 GB.
Nemotron 3.5 Super VL combines a hybrid Mamba Transformer architecture with a Mixture of Experts (MoE) design, featuring 120 billion total parameters while activating approximately 12 billion per token. The multimodal model supports reasoning, coding, tool execution, and understanding of images, documents, charts, and videos. NVIDIA reports training the model on approximately 26 trillion tokens, followed by additional training using 3.8 million samples and reinforcement learning.
NVIDIA claims Nemotron 3.5 Super VL delivers up to 5 times higher inference throughput than competing open models in its class, based on selected agentic benchmarks. The company also describes it as offering leading accuracy for specialized enterprise tasks. These results reflect NVIDIA's published evaluations and should not be interpreted as a universal performance advantage across different workloads, hardware configurations, or competing models.
NVIDIA Nemotron 3.5 Model Comparison
| Specification | Super VL | Bolt 74B-A9B 24GB | Bolt E4B 4GB |
|---|---|---|---|
| Total Parameters | 120B | 74B | Not specified |
| Active Parameters | 12B | 9B | Not specified |
| Memory Footprint | Depends on precision | 24 GB target | 4 GB target |
| Optimization | MoE, NVFP4 available | Primarily 2 bit quantization | Compact edge inference |
| Target Applications | Enterprise multimodal agents | Local reasoning and coding | Voice and interactive agents |
| Availability | Available | Available | October 23, 2026 |
Specifications reflect NVIDIA's published documentation. The 24 GB and 4 GB figures are targeted model footprints, not guaranteed total system memory requirements.
Nemotron 3.5 Bolt 74B-A9B 24GB uses structured compression and aggressive quantization to reduce the larger architecture into a model designed for 24 GB of memory. In its technical research, NVIDIA reports approximately 84 tokens per second using llama.cpp and speculative decoding on an RTX Spark system. The result was obtained under specific benchmark conditions and does not establish equivalent performance across consumer GeForce GPUs.
NVIDIA also reports that Bolt 74B-A9B performs competitively against several open models with similar memory requirements on agentic and coding benchmarks. However, independently reproduced comparisons will be important for determining how its aggressive quantization affects reasoning accuracy, responsiveness, and practical task completion.
Meanwhile, Nemotron 3.5 Bolt E4B 4GB targets conversational agents, voice assistants, interactive game characters, and embedded applications running on GeForce RTX, Jetson, and RTX Spark hardware. Its smaller footprint is intended to make local AI deployment more accessible, although NVIDIA has not yet published comprehensive independent performance results for the upcoming model.
Super VL and Bolt 74B-A9B are available through NVIDIA's model distribution channels, with support for frameworks including Ollama, LM Studio, llama.cpp, and vLLM. The announcement continues NVIDIA's expansion of local AI models across RTX and DGX hardware, offering developers additional deployment options without depending exclusively on cloud infrastructure.
NVIDIA's latest Nemotron models highlight how AI development is increasingly focused on performance per gigabyte rather than parameter counts alone. Compressing a 74B model into a 24 GB target footprint could make sophisticated local AI applications more practical for users with compatible RTX hardware.
The smaller Bolt E4B model is particularly interesting for gaming applications, where lightweight conversational AI could eventually support more interactive NPCs and responsive in game assistants.
However, memory efficiency and benchmark performance are not the same as real world reliability. NVIDIA's reported advantages remain dependent on test conditions, inference frameworks, and model quantization. Independent testing will be essential to establish whether these models offer meaningful improvements over existing alternatives.
Would you use your GeForce RTX gaming PC to run local AI agents with 24 GB or even 4 GB memory requirements, or would you prefer dedicated AI hardware?
