Nvidia Documents 121B-Parameter Nemotron 3.5 Super VL Model to Target Claude and GPT

Nvidia has introduced Nemotron 3.5 Super VL, an open vision-language model featuring roughly 121 billion total parameters and a 1 million-token context window. The architecture uses a hybrid mixture-of-experts backbone called NemotronH_Omni_Reasoning_V3, fielding 512 routed experts where only 22 activate per token to reduce inference costs. For visual processing, the model integrates a RADIO v2.5-H encoder to handle images alongside text inputs. Nvidia claims the model outperforms Anthropic’s Claude Opus 5 and OpenAI’s GPT-6.1 Sol on specialized benchmarks after post-training, though these metrics currently lack independent verification. The model incorporates Mamba-Transformer efficiencies and multi-token prediction to accelerate output generation for targeted agentic tasks.