Multiverse Computing Launches Comet 60B
aiinsider ·
Multiverse Computing has launched Comet 60B, an open-source multimodal AI model that reduces the parameter count of Nvidia’s Nemotron 3.5 Super VL from 120 billion to 60 billion. Despite halving the total parameters, the new model retains 12 billion active parameters and continues to support text, image, and video processing. Built using Multiverse’s CompactifAI compression framework, the model is designed to require less GPU memory while maintaining performance levels within 5% of the original Nvidia model across most benchmarks. In several specific evaluations, including tau2-Bench telecom, BabyVision, and AutomationBench-AA, Comet 60B actually outperformed its larger predecessor. The architecture utilizes a mixture-of-experts approach, allowing it to activate only necessary parts of the model for each task. Available on Hugging Face in BF16, FP8, and NVFP4 formats, the release targets developers building agentic AI systems for complex multi-step tasks. This launch follows a similar strategy with Pulsar 16B, leveraging early access to Nvidia models to create efficient compressed alternatives.
aiinsider