Nvidia releases Nemotron 3.5 Super VL with open weights, training data and reinforcement-learning environments

Nvidia released Nemotron 3.5 Super VL on October 15. It is an open-weight multimodal model built for agents that run continuously inside enterprise systems, according to The Decoder. The model is a mixture-of-experts network with 120 billion total parameters and 12 billion active parameters. It reads text, images and video clips of up to two minutes, as the Russian-language channel AI ML Big Data summarized. The context window reaches one million tokens, and Nvidia licenses the model under OpenMDW-1.1.

This release includes more than most open-weight launches. Nvidia published a base checkpoint, a post-trained checkpoint and a 4-bit compressed checkpoint. It also released the datasets, the training recipes and the reinforcement-learning environments, and made the model available on Hugging Face, OpenRouter, ModelScope and its own NIM service, The Decoder reported. Launch coverage describes a hybrid architecture that combines Mamba-2, mixture-of-experts and attention layers, with a C-RADIO vision encoder and 512 routed experts per expert layer. The state-space layers make the million-token context feasible, because they limit growth in the key-value cache, the memory that stores earlier tokens during generation (also called the KV-cache), compared with what a pure-attention architecture of the same depth would require.

On serving speed, Nvidia says the model generates more than 600 tokens per second on a single GPU, up to five times as fast as open competitors in its class. It credits built-in multi-token prediction, a speculative-decoding method that speeds generation on DGX Spark by up to 2.3 times, and NVFP4 quantization with distillation that Nvidia says keeps BF16 accuracy, according to Serverflow. These are vendor figures.

Independent verification is more limited. Artificial Analysis scores the model at 26 on its Intelligence Index, well above the average for open-weight models of similar size. It also reports roughly twice the predecessor's index score with about 60% fewer output tokens per task. The stronger claim comes from VentureBeat, which reports that after post-training, the model beats Anthropic's Claude Opus 5 and OpenAI's GPT-6.1 Sol on specialized task benchmarks. That comparison pits a domain-tuned model against general-purpose ones, and OrcaRouter noted that no independent scoreboard existed at launch. The weights, the license and the architecture can be verified now. The head-to-head wins over closed frontier models are Nvidia's claim until an outside group reproduces them on tasks Nvidia did not select.