61.1B total / 12.3B active (marketed as 60B total / 12B active MoE)
262,144 tokens evaluated (configuration permits 1,048,576 positions)
NVIDIA Nemotron Open Model License
Open-weight multimodal reasoning and agent model produced by compressing NVIDIA Nemotron 3.5 Super VL with Multiverse Computing's quantum-inspired CompactifAI framework. Comet reduces the parent model from 124.5B total parameters to 61.1B while retaining approximately the same active compute at 12.3B parameters per token. It preserves the 88-layer hybrid Mamba-2, Transformer, latent Mixture-of-Experts, attention, RADIO ViT-H/16 vision, and Multi-Token Prediction architecture, while reducing the number of routed experts per MoE layer from 512 to 256 and continuing to activate 22 experts plus one shared expert. Comet accepts text, images, and video and produces text, with configurable thinking, non-thinking, tool calling, multimodal document understanding, coding, and agentic operation; audio is not supported. After tensor-network profiling and structural compression, Multiverse healed the model through token-level knowledge distillation from Nemotron 3.5 Super using single-turn reasoning and code, multi-turn terminal and tool-use trajectories, images, and video. A final GRPO-style stage targeted multi-turn interaction, tool use, hallucination reduction, reasoning, and verbosity. The family includes a 122.3GB BF16 checkpoint requiring two H200s for the full evaluated 256K context, a 66.3GB FP8 checkpoint that fits on one H200, and a 40.5GB mixed NVFP4/FP8/BF16 checkpoint for Blackwell hardware. The released configuration permits 1M positions, but Multiverse reports evaluation only through 262,144 tokens. Multiverse reports BF16 scores of 88.64 on GPQA Diamond, 70.00 on SWE-bench Verified, 50.93 on SWE-bench Pro, 80.79 on tau2-Bench, 33.18 on AutomationBench-AA, 67.40 on MMMU-Pro, and 73.26 on Video-MME with subtitles. The FP8 checkpoint scores 71.00 on SWE-bench Verified, 51.48 on SWE-bench Pro, 81.01 on tau2-Bench, and 30.93 on BabyVision. In Multiverse's H200 serving tests, BF16 on four GPUs delivered 20% more chat and 23% more agentic output tokens per second per GPU than Nemotron 3.5 Super, while FP8 on two GPUs delivered 87% and 78% more respectively.