Insider Brief
- Multiverse Computing has launched Comet 60B, an open-source multimodal AI model that cuts Nvidia Nemotron 3.5 Super VL from 120 billion to 60 billion total parameters while retaining text, image and video support and the same 12 billion active parameters.
- The company said the smaller model requires less GPU memory and remains within 5% of the original model’s accuracy across most tested benchmarks, while outperforming it on some evaluations including tau2-Bench telecom, BabyVision and AutomationBench-AA.
- Comet 60B was built with Multiverse’s CompactifAI compression framework and is available on Hugging Face in BF16, FP8 and NVFP4 formats.
Multiverse Computing has launched Comet 60B, an open-source multimodal AI model that cuts the total parameter count of Nvidia’s Nemotron 3.5 Super VL in half while retaining support for text, images and video.
According to the San Sebastián, Spain-based company, Comet 60B has 60 billion total parameters, compared with 120 billion for the original Nvidia model, while keeping the same 12 billion active parameters. Both use a mixture-of-experts architecture, which activates only part of the model for each task.
Multiverse claims the lower total parameter count reduces the amount of GPU memory needed to load the model, allowing it to run on less hardware. Because the active parameter count remains unchanged, the amount of computation used for each generated token stays at the same level, according to the company.
The model is designed for developers building agentic AI systems that work across text, images and video, including coding, document analysis and other multi-step tasks. Multiverse said Comet 60B stays within 5% of the original model’s accuracy across most of the benchmarks it tested and exceeds it on several individual evaluations.
Specs
Key specifications include:
- Base model: Nvidia Nemotron 3.5 Super VL, with 120 billion total and 12 billion active parameters.
- Comet 60B: 60 billion total and 12 billion active parameters.
- Architecture: Mixture of experts.
- Modalities: Text, image and video.
- Formats: BF16, FP8 and NVFP4.
- Performance: Within 5% of the original model across most tested benchmarks, according to Multiverse.
Multiverse produced Comet 60B using CompactifAI, its model-compression framework. The process begins by analyzing the model with tensor-network techniques to identify redundant or lower-impact information. The company then creates compressed versions by removing selected parts of the model, followed by a retraining stage intended to recover capabilities lost during compression.
The retraining then uses data from single-turn text tasks, multi-turn agent interactions, images and video and a final reinforcement-learning stage is used to improve areas including multi-turn conversations, reasoning and reducing incorrect outputs.
Multiverse reported that Comet 60B outperformed the original Nvidia model on several tests, including the telecom portion of tau2-Bench, where it scored 83.11 compared with 81.91. It also posted higher results on BabyVision and AutomationBench-AA.
On other benchmarks, the smaller model trailed the original by narrower margins. Comet scored 88.57 versus 90.40 on GPQA Diamond, 51.48 versus 52.44 on SWE-bench Pro and 72.44 versus 74.80 on Video-MME, according to the company.
Multiverse pointed out that it received early access to Nvidia’s model, following a similar arrangement for Pulsar 16B, which it launched in June based on a compressed version of Nvidia Nemotron 3 Nano. Comet 60B is available now through Hugging Face in BF16, FP8 and NVFP4 versions, along with technical documentation covering the compression methodology and benchmark results.
