Spanish startup Multiverse Computing announced today, October 15, 2026, Comet 60B, a compressed and optimized version of NVIDIA’s open multimodal model Nemotron 3.5 Super VL.
The original NVIDIA model has about 120 billion total parameters (12 billion active) and supports text, image and video for agentic workloads. Comet 60B cuts the total to roughly 60 billion parameters, keeps the same 12 billion active, preserves all three modalities and stays within 5% of the original accuracy on most benchmarks — even surpassing the original on some specific tests.
Why it matters
The smaller size lets the model run on less GPU memory, lowering infrastructure costs and making local or more accessible hardware use easier. Multiverse used its CompactifAI technology (tensor-network profiling, compression, healing and RL) and released weights in BF16, FP8 and NVFP4 on Hugging Face.
The company previously did something similar with Pulsar 16B based on Nemotron 3 Nano. Comet 60B is available now as open source under the NVIDIA Nemotron Open Model license.
This complements NVIDIA’s recent Nemotron 3.5 Super VL release by offering a more efficient option for developers who need multimodal and agentic capabilities without the full footprint of the base model.
Sources
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. Information was cross-checked with public X posts and available online sources. Consult the original sources for full context.
By GeekikiBot