Multiverse Computing Slashes Nemotron 3.5 Super VL to 60B Parameters, Maintaining Multimodal Performance

  • Multiverse Computing launched Comet 60B, a compressed version of NVIDIA's Nemotron 3.5 Super VL, reducing parameters from 120B to 60B while maintaining 95% accuracy across text, image, and video tasks.
  • Comet 60B retains the original model's 12B active parameters, ensuring similar per-token compute efficiency but requiring less GPU memory for deployment.
  • The model outperforms the original in specific benchmarks, including telecom agent tasks (83.11 vs 76.75) and visual reasoning (30.93 vs 26.80).
  • Comet 60B is available open-source on Hugging Face in BF16, FP8, and NVFP4 formats.
  • Multiverse Computing developed Comet 60B using its CompactifAI framework, which includes tensor-network profiling, compression, healing, and reinforcement learning.

Multiverse Computing's Comet 60B exemplifies the growing emphasis on AI model efficiency, particularly for multimodal and agentic applications. As enterprises demand more cost-effective deployments, compression technologies like CompactifAI could become a key differentiator. The open-sourcing of Comet 60B also signals a broader industry shift toward democratizing access to advanced AI capabilities, potentially accelerating innovation in niche use cases.

Adoption Pace
How quickly enterprises will deploy Comet 60B for agentic workloads, given its cost and efficiency advantages.
Competitive Response
Whether NVIDIA or other AI model providers will accelerate their own compression efforts to match Multiverse's efficiency gains.
Benchmark Validation
The extent to which independent evaluations confirm Comet 60B's claimed performance parity with the original model.