Multiverse Computing debuts Comet 60B, touting multimodal performance built on NVIDIA Nemotron technology

Multiverse Computing on Oct. 15, 2026 announced the launch of Comet 60B, positioning it as a multimodal large language model designed to handle multiple types of input such as text and other data modalities. The announcement emphasizes efficiency, saying Comet 60B is a compressed and optimized variant of an NVIDIA Nemotron 3.5 “Super VL” model, with the parameters reduced to about half the size.

According to the release carried by Yahoo Finance, the shift to fewer parameters is intended to retain multimodal capability while lowering the model’s size. That framing points to a broader industry focus on getting stronger results without the full compute and storage demands typically associated with larger model deployments.

The company’s update also ties Comet 60B’s rollout directly to NVIDIA, describing the model launch as happening in collaboration with NVIDIA. It does not, in the details provided here, specify the exact engineering process, the training dataset strategy, or the hardware configuration used for training and inference.

Comet 60B is described as “optimized” in addition to being compressed, a distinction that usually implies more than simply pruning parameters, but the announcement coverage in the packet does not lay out which optimization techniques were used. It also does not provide benchmarks, latency figures, cost-per-inference estimates, or comparisons against competing multimodal models.

For NVIDIA, the announcement is part of a continued effort to extend its AI platform beyond raw model development into ecosystem adoption. Nemotron refers to NVIDIA’s multimodal model line, and Super VL in this context appears to be the “super visual-language” variant label used in the release. In practice, partnerships like this can help NVIDIA’s software and model tooling become embedded in third-party offerings that target enterprise and developer use cases.

For the market, the claim of “half the parameters” matters because parameter count is closely watched as a proxy for how expensive it can be to host and run frontier-style models. If the performance holds up, smaller models can be easier to deploy in production settings and can reduce bottlenecks in scaling services that depend on fast, multimodal reasoning.

Still, several key details remain undisclosed in the information available here. The announcement coverage does not include any public model weights or licensing terms, does not name a specific inference framework, and does not report measurable results such as accuracy on standardized multimodal benchmarks or head-to-head comparisons against the underlying NVIDIA Nemotron 3.5 Super VL baseline.

What to watch next is whether Multiverse Computing publishes benchmark results, deployment guidance, and a clearer description of how Comet 60B’s compression approach translates into real-world throughput and cost. Developers and enterprise buyers are likely to look for evidence that the model’s quality is maintained as parameter counts drop, and for clarity on how the NVIDIA collaboration is reflected in tooling and supported runtimes.