Comet 60B: Multiverse Computing squeezes NVIDIA's Nemotron 3.5 Super VL into half the parameters

Multiverse Computing launched Comet 60B on Thursday, a compressed version of NVIDIA's brand-new Nemotron 3.5 Super VL: half the parameters at 60B total, all 12B active kept, accuracy within 5% of the original — and it beats the 120B model on some benchmarks.

Half the model, nearly all the accuracy

Multiverse Computing, the San Sebastián AI company, announced Comet 60B on Thursday via a GlobeNewswire release: a compressed and optimized version of NVIDIA's Nemotron 3.5 Super VL, the 120-billion-parameter multimodal mixture-of-experts model that launched the same day. The compression is dramatic — total parameter count falls from 120 billion to 60 billion while the active parameter count stays at 12 billion and all three modalities, text, image and video, are retained.

The pitch is aimed squarely at agent deployments. Total parameter count is what decides how much GPU memory a model needs just to load, so halving it means Comet 60B fits on meaningfully less hardware — while per-token compute is unchanged because the same 12 billion active parameters do the work each step.

Multiverse says Comet 60B stays within 5% of the original's accuracy across text, image and video tasks — and actually improves on it in some areas. The reported gains: τ²-Bench telecom (83.11 vs. 81.91), BabyVision (30.93 vs. 26.80) and AutomationBench-AA (33.38 vs. 31.16). On harder reasoning it nearly matches the original: GPQA Diamond (88.57 vs. 90.40), SWE-bench Pro (51.48 vs. 52.44) and τ²-Bench (81.01 vs. 81.48). Visual tasks track closely too: CharXiv reasoning 72.90 vs. 73.70, MathVision 69.34 vs. 70.20, OCRBench v2 66.28 vs. 67.30, MMLongBench-Doc 58.51 vs. 60.00, MMMU-Pro 66.88 vs. 70.35 and Video-MME 72.44 vs. 74.80. It also scores above the previous-generation NVIDIA Nemotron 3 Super on all eight benchmarks shown.

How it was built: CompactifAI's four steps

Comet 60B comes out of Multiverse's CompactifAI framework, which runs in four stages: tensor-network profiling to identify redundant or low-impact parts of the model without changing it; compression, which generates multiple reduced candidates and removes the least relevant parts with minimal accuracy loss; healing, which recovers the base model's capabilities and knowledge using single-turn text (math, science, reasoning), multi-turn agentic trajectories and image/video data; and a final GRPO-style reinforcement-learning stage targeting multi-turn conversation, hallucination mitigation and chain-of-thought reasoning.

On top of that, the model ships quantized in BF16, FP8 and NVFP4, so teams can match precision to their hardware and latency targets.

The pattern: a launch-day compression specialist

This is Multiverse's second day-one NVIDIA compression. The company received early access to Nemotron 3.5 Super VL — and did the same for NVIDIA's Nemotron 3 Nano, releasing the compressed Pulsar 16B in June 2026. Comet 60B is available today on Hugging Face in all three precisions, with full methodology and per-benchmark documentation alongside. Multiverse — headquartered in Donostia-San Sebastián with offices in the US, Canada and Europe, and more than 100 customers including Iberdrola, Bosch and the Bank of Canada — frames it as making advanced AI deployable inside organizations' own infrastructure.

Why it matters

Same-day launches are rare; same-day compression is the interesting part. NVIDIA's pitch for Nemotron 3.5 Super VL is specialized, always-on agents that enterprises post-train on their own data — and the bottleneck for that vision isn't intelligence, it's hardware footprint and inference cost. A 60B-weight drop-in that keeps the 12B-active compute budget and stays within 5% accuracy, while actually winning a few agentic benchmarks, turns a data-center-heavy deployment into a much cheaper one to serve. The caveat, as always with vendor-published numbers: these are Multiverse's benchmarks on Multiverse's chosen evaluations. Independent leaderboard scores will tell whether "within 5%" holds in the wild.

Sources: Multiverse Computing's official announcement, via GlobeNewswire (Oct 15, 2026); Tech Insider's Nemotron 3.5 Super VL launch coverage (Oct 15, 2026).