NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 released

Nemotron 3.5 Super VL is NVIDIA's new open vision-language model: it takes text, images and videos of up to two minutes and answers in text, and it targets reasoning, coding, tool use and document understanding. It has 120 billion parameters but is a mixture-of-experts model, so only about 12 billion are used for each token, and NVIDIA says its mix of Mamba-2, expert and attention layers is there to support long-context processing, up to 1 million tokens. A built-in multi-token prediction head predicts upcoming tokens for the model to check, which lets it generate faster through speculative decoding.

Several checkpoints arrived together, and this BF16 one is the finished model at full 16-bit precision. NVIDIA says it is meant for customization, adapting to domain data, building quantized variants and full-precision evaluation, and the card lists four H100 GPUs or two GB200s for deployment. For efficient serving the card points to the 4-bit NVFP4 version, which was distilled from these weights; for training from an earlier stage, before fine-tuning and reinforcement learning, there is a separate Base checkpoint.

NVIDIA's own benchmark table, collected with its NeMo Gym tooling, shows a mixed picture against other models. Compared with Qwen3.5-122B-A10B, Super VL scores higher on the BrowseComp web-research test (78 vs 58.1) and Terminal-Bench 2.1 (60.67 vs 48.70), but lower on MMLU-ProX (76.91 vs 83.2) and the tau2-bench telecom agent test (73.19 vs 87.2). Other models in the same table, including Qwen3.8-Flash-Next, DeepSeek-V4-Flash and GLM-5.3-Flash, score higher than Super VL on most rows. The model card links NeMo Gym recipes for reproducing the evaluation settings.