New Multimodal AI Model Reduces Memory Requirements
Multiverse Computing's Comet 60B halves parameter counts while matching the accuracy of its base architecture within 5 percent.
Updated on Oct. 15, 2026 in Artificial Intelligence
Live Poll
Would your workplace consider switching to smaller, more efficient AI models to reduce operating costs?
Multiverse Computing has launched Comet 60B, a multimodal AI model supporting text, image, and video. This released model is based on the 120-billion-parameter NVIDIA Nemotron 3.5 Super VL but is compressed to 60 billion total parameters.
Why it matters
The model aims to reduce infrastructure costs and GPU memory demands for agentic AI tasks. By shrinking base model requirements while retaining most of their performance, developers can deploy complex workloads on more accessible hardware.
Comet 60B matches the original Nemotron 3.5 Super VL accuracy within 5 percent across all tested benchmarks. It supports BF16, FP8, and NVFP4 formats, enabling deployment with significantly lower memory footprints than its 120-billion-parameter ancestor.
The players
Multiverse Computing
A San Sebastian-based firm specializing in tensor-network methods and AI model compression for industrial applications.
NVIDIA
The semiconductor manufacturer whose Nemotron 3.5 Super VL architecture serves as the foundation for the Comet 60B model.
The details
Multiverse Computing utilizes its CompactifAI framework to process the base model through tensor-network profiling, compression, and healing. This technique identifies redundant internal structures within the neural network, pruning them while using reinforcement learning to maintain original output quality. The final training stage incorporates Group Relative Policy Optimization, an algorithm that improves model alignment during the compression transition.
Timeline
June 2026: Multiverse Computing released the Pulsar 16B model.
October 15, 2026: The company launched the Comet 60B model.
The Tech Race
This release follows the company’s trend of optimizing high-parameter models, building on the earlier Pulsar 16B release. By targeting memory efficiency, the firm is competing in a crowded field of companies aiming to bring heavy agentic AI capabilities to resource-constrained environments.
Developers and enterprises can now deploy multimodal AI capabilities that previously required 120-billion-parameter hardware on reduced infrastructure footprints. This transition allows teams currently using heavy models to potentially lower their cloud compute and memory costs starting immediately.
The takeaway
Comet 60B demonstrates that compression techniques can maintain high performance while cutting parameter counts by half. Watch for how these memory-saving models perform in live agentic workflows versus non-compressed alternatives.
Further reading
For broader trends in model efficiency, visit the Artificial Intelligence section.
Live Poll
Would your workplace consider switching to smaller, more efficient AI models to reduce operating costs?
