Multiverse Computing Launches Comet 60B in collaboration with NVIDIA: Multimodal Capabilities at Half the Parameters

SAN SEBASTIAN, Spain, Oct. 15, 2026 (GLOBE NEWSWIRE) — Multiverse Computing today announced Comet 60B, a compressed and optimized version of NVIDIA’s Nemotron 3.5 Super VL. The original model is a multimodal Mixture-of-Experts (MoE) model with 120 billion total parameters and 12 billion active parameters, built for long, complex tasks and for agentic workloads that combine text, image and video. Comet 60B halves the total parameter count to 60 billion, keeps all three modalities and the same 12 billion active parameters, and improves the original model in some specific areas and benchmarks, while for the rest stays within 5% of the original model’s accuracy across text, image and video tasks.

Multiverse Computing received early access to the model, as it did when it launched Pulsar 16B in June 2026, built on a compressed and optimized version of NVIDIA Nemotron 3 Nano.

For teams deploying agents, total parameter count and more decides how much GPU memory a model needs just to load. Cutting total parameters in half means Comet 60B fits on less hardware than the original, while the unchanged active parameter count keeps per-token compute at the same level.

“Agents need models that can read documents, understand images and follow video across long tasks, and they need to run where the data lives,” said Enrique Lizaso, co-founder and CEO of Multiverse Computing. “Nemotron 3.5 Super VL is an excellent multimodal model. With Comet 60B, we have cut its size in half while keeping accuracy within 5% of the original, so companies can run it on less hardware and at lower cost. That is exactly what our compression technology is built for.”

Comet 60B, Multiverse Computing’s new open-source multimodal model, is designed for teams building and running agentic AI workloads. Combining strong coding and agent capabilities with support for text, images and video, it gives developers a versatile foundation for creating assistants, automating complex tasks and building applications that work across different types of content.

With 50% fewer parameters, Comet 60B reflects the company’s commitment to making advanced AI more efficient and accessible. Available to the community, it enables developers and organisations to explore new multimodal use cases and build AI solutions around their own workflows.

How Comet 60B is built

Comet 60B is built on top of a compressed and optimized version of NVIDIA Nemotron 3.5 Super VL. Multiverse Computing produced Comet 60B through its framework CompactifAI:

Tensor-network profiling. CompactifAI starts by analyzing the model with techniques based on tensor networks. This analysis identifies which parts of the model carry redundant or low-impact information, without changing the model itself at this stage.

Compression. Based on the profiling results, multiple candidates are generated, this step removes the least relevant parts of the model with minimal impact on accuracy.

Healing. After compression, the reduced model is healed to recover the original capabilities and knowledge from the base model. We used data from three areas: single-turn text (math, science, reasoning, etc), multi-turn agentic trajectories (terminal, tool-use, etc), and images and videos.

Reinforcement Learning. A final GRPO-style training (Group Relative Policy Optimization) improves key areas like multi-turn conversation, mitigation of hallucination, chain of thought and reasoning.

As a final additional step, on top of CompactifAI framework, we also quantize the model to ship in BF16, FP8 and NVFP4 formats, so customers can choose the precision that matches their hardware and latency targets.

Key specifications

• Base model: NVIDIA Nemotron 3.5 Super VL (120B total, 12B active)
• Comet 60B: 60B total parameters, 12B active, MoE architecture
• Modalities: text, image and video
• Accuracy: within 5% of the original model, improved in some benchmarks
• Formats: BF16, FP8, NVFP4

Benchmarks where Comet 60B outperforms NVIDIA Nemotron 3.5 Super VL: even at half the size, our model is able to recover and slightly improve in some of the benchmarks such as the ones showed above. The post-training stage, specifically the RL part, allow us to improve the original model. The largest improvements are tau2-Bench telecom (76.75 vs 83.11), BabyVision (26.80 vs 30.93) and AutomationBench-AA (31.16 vs 33.38).

Text and agentic tasks: these benchmarks measure how well a model reasons, writes code and acts as an agent. Comet 60B nearly matches the original on GPQA Diamond, a set of graduate-level science questions (88.57 vs 90.40), SWE-bench Pro, which asks the model to resolve real software engineering issues (51.48 vs 52.44) and τ²-Bench, which tests agents that use tools to complete customer service tasks (81.01 vs 81.48). It also scores above the previous-generation NVIDIA Nemotron 3 Super on all eight benchmarks shown.

Image and video benchmarks: these tests check how well a model understands visual content. Results stay very close to the original on CharXiv reasoning, which tests reasoning over scientific charts (72.90 vs 73.70), MathVision, visual mathematical reasoning (69.34 vs 70.20), OCRBench v2, text recognition in images (66.28 vs 67.30), MMLongBench-Doc, understanding of long documents (58.51 vs 60.00), MMMU-Pro, multidisciplinary image-based reasoning (66.88 vs 70.35) and Video-MME, video understanding (72.44 vs 74.80).

τ²-Bench by domain: the same agent benchmark split by scenario, where the model must follow policies and use tools to help a simulated customer. Comet 60B reaches 82.68 on retail (original 82.79), 77.25 on airline (79.75) and 83.11 on telecom, where it scores above the original (81.91).

Across text, agentic, image and video benchmarks, Comet 60B delivers results almost identical to the original NVIDIA Nemotron 3.5 Super VL with half the parameters: 60 billion instead of 120 billion. The reduction in size is important for cost. Fewer parameters mean less GPU memory is needed to load and run the model, so companies can deploy it on less hardware and lower their infrastructure and inference costs, without giving up the capabilities of the original model.

Availability
Comet 60B is available today on Hugging Face as BF16, FP8 and NVFP4. Full technical documentation, including methodology, evaluation setup, and per-benchmark results, is available here. To explore Multiverse Computing’s open-source model releases, visit huggingface.com/MultiverseComputingCAI.

About Multiverse Computing
Multiverse Computing is a leader in sovereign and efficient AI. The company develops fast, efficient, and highly specialized AI models that enable organizations to deploy advanced artificial intelligence securely within their own infrastructure, ensuring full control over data, governance, and compliance. Headquartered in Donostia-San Sebastián, Spain, with offices in the United States, Canada, and across Europe, Multiverse serves more than 100 global customers, including Iberdrola, Bosch, and the Bank of Canada. For more information, visit multiversecomputing.com

Media contact
LaunchSquad for Multiverse Computing

multiverse@launchsquad.com

Photo accompanying this announcement are available at:

https://www.globenewswire.com/NewsRoom/AttachmentNg/9402d27f-1017-410d-b6d6-71bb12abde03

https://www.globenewswire.com/NewsRoom/AttachmentNg/0f7d688e-2e00-434b-95ae-e0b580fa7e9f

https://www.globenewswire.com/NewsRoom/AttachmentNg/f7f96aa9-1d3c-4177-a12f-8fb6692139f1

https://www.globenewswire.com/NewsRoom/AttachmentNg/2088f741-a9f9-4d80-8d14-46d2862629e6

https://www.globenewswire.com/NewsRoom/AttachmentNg/da029a7d-e2a1-40fa-96e6-c69f1e2f60cd