NVIDIA released the open-weight Nemotron 3.5 Transcribe model, which supports transcription in 40 languages, translation in 31 languages, and speaker attribution in 13 languages, ranking top on Hugging Face Open ASR and other leaderboards. It also open-sourced the Granary v2 multilingual dataset and fine-tuning tools, with a medical fine-tuning case reducing word error rate by 45.8%.
NVIDIA and AI4Bharat at IIT Madras have released Nemotron-pretraining-Indic-v1, a dataset covering 22 Indian languages with about 2.8 trillion tokens, along with an open, GPU-accelerated data generation pipeline featuring prompts, LLM workflows, and quality checks for reuse.
NVIDIA launched the open multimodal model Nemotron 3.5 Super VL, built for continuous, multi-step agent tasks. It uses a MoE architecture with 120 billion total parameters and 12 billion active parameters, delivering up to 5x single-GPU throughput on agent workloads versus similarly sized open models. Base model, post-trained model, NVFP4 quantized weights, training recipes, datasets and reinforcement learning environments are also open-sourced.
Nvidia open-sourced Nemotron-Labs-TwinTower diffusion language model, which uses a "twin tower" architecture to overcome the serial decoding bottleneck of autor