Hugging Face AI Updates: August 11, 2026

1. NVIDIA Ships Open-Weights Magpie Multilingual TTS For Low-Latency Voice Agents

Hugging Face published NVIDIA’s release of Magpie Multilingual TTS, a 364M-parameter open-weights text-to-speech model aimed at low-latency voice agents, distributed on the Hub under the NVIDIA Open Model License. The model now covers 12 languages, including English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese, with male and female speakers each, and reports time-to-first-audio as low as 32 ms on a B200 GPU. Practitioners can fine-tune the open checkpoints via NeMo or deploy through NVIDIA NIM containers, giving teams full weights and deployment control for real-time conversational systems. Source

2. Multiverse Computing Open-Sources Cheaper Knowledge Distillation At Scale

Hugging Face hosted a technical write-up from Multiverse Computing on cutting the cost of knowledge distillation, along with an open-source implementation and a distilled model on the Hub. The approach combines offline distillation, which caches the teacher model’s top-100 logits so both models need not sit in memory at once, with a fused chunked KL loss that processes the vocabulary in chunks to avoid memory spikes, reportedly dropping peak VRAM from about 250GB to 128GB at 32K context and cutting setup from four GPU nodes to one. The team released a 3.2B-parameter student model distilled from Llama 3.1 8B, lowering the barrier for practitioners who want to compress large models on limited hardware. Source