NVIDIA releases Nemotron 3.5 Super VL, a 120B open multimodal MoE, plus Bolt edge models

NVIDIA's new open model targets always-on enterprise agents: reading long documents, running OCR, interpreting images and video, and cross-checking information across sources in one request. It has 120B total parameters with 12B active, and NVIDIA claims up to 5x the single-GPU throughput of comparable models.

NVIDIA released open datasets on Hugging Face alongside it, covering tool use, financial reasoning and 45 vision-language sets. Red Hat shipped optimized FP8 and speculative-decoding checkpoints. For edge use there is the 24GB Nemotron 3.5 Bolt 74B-A9B, with a 4GB Bolt E4B still to come.

The release adds to the week's flood of open weights, alongside Meta's Muse Spark Open, Mistral Large 4 and StepFun's Step 5. Axios notes that hyperscaler custom chips such as Microsoft's Maia are turning NVIDIA customers into rivals. Open models that sell NVIDIA hardware are part of NVIDIA's response.

Watch for independent multimodal benchmarks and how the Bolt checkpoints perform on consumer GPUs.