Multiverse Computing releases Comet 60B, a half-size version of NVIDIA's Nemotron model

The story so far, in brief; the legwork, and the full article, stay with the original outlet.

Multiverse Computing, the San Sebastián company whose business is making AI models smaller, announced Comet 60B on 15 October 2026. It is a compressed version of NVIDIA's Nemotron 3.5 Super VL, a multimodal mixture-of-experts model with 120 billion parameters in total, 12 billion of them active. Comet keeps 60 billion, the same 12 billion active, and all three of text, image and video. Multiverse says accuracy is no more than 5% below the original's, and better on some benchmarks.

The arithmetic is the selling point. The total count decides how much GPU memory you need merely to load the thing, so halving it means fewer chips, while the active count, and with it the compute per token, stays where it was. Multiverse got early access to the model, as it did for Pulsar 16B in June, which was built on Nemotron 3 Nano. "We have cut its size in half while keeping accuracy within 5% of the original," said Enrique Lizaso, co-founder and CEO.

The model is open source and on Hugging Face in BF16, FP8 and NVFP4. It came out of the company's CompactifAI framework: tensor-network analysis to find the parts that matter least, removing them, retraining to recover what was lost, then a round of reinforcement learning. Multiverse counts more than 100 customers, Bosch, Iberdrola and the Bank of Canada among them.