tensorfolk

Google DeepMind unveils Genie 4, a real-time world model that also generates sound and dialogue

Big news

DeepMind's Genie 4 builds explorable worlds you steer with the keyboard and edit by text mid-session, such as changing the weather or adding people. It is the lab's first real-time world model to generate audio natively, including music and dialogue, with more realistic people and physics. It is a research preview for trusted testers, with Google AI Ultra and select Gemini API partners next.

Why it matters

Interactive world models are edging toward usable simulators for games, agent training and robotics, though sessions still drift over time.

Source: Read the full story on Google DeepMind

world-modelsdeepmindgeniesimulationaudio

More in Generative Media

Generative Media / Video Generation

xAI posts the first seven minutes of its Grok Imagine-made Odyssey series

xAI's Grok Imagine account released Episode 1 of an AI-generated Odyssey series, about seven minutes long, made with Wonder Studios and director Dev Walker. It is the first public piece of the full-length film Musk promised in July. Early reviews say the footage holds together better than expected but lacks artistry.

Why it matters: A real test of whether current text-to-video tools can sustain multi-minute narrative work, not just short clips.

Read on TechCrunch
Generative Media / Video Generation

Kyutai open-sources CADRE causal video autoencoders for real-time world models

Kyutai released seven causal video autoencoders, where each frame depends only on past frames, along with 1.2B-parameter world models built on them. The world models reach about 20 fps on a single H100 with action lag under a quarter second. Code is MIT-licensed and weights are non-commercial.

Why it matters: Targets the latency bottleneck that blocks playable, interactive video world models.

Read on Kyutai
Generative Media / Voice, Speech & Voice Agents

NVIDIA opens Nemotron 3.5 Transcribe, a 30B-A4B speech model for 40 languages, plus 463k-hour Granary v2

NVIDIA published open weights for Nemotron 3.5 Transcribe, a mixture-of-experts speech model with about 4B active parameters. One model handles transcription in 40 languages, translation to and from English for 31 of them, and labeling of up to eight speakers in hour-long recordings. Alongside it comes Granary v2, a speech corpus covering 48 languages.

Why it matters: One commercially usable open model replaces separate ASR, translation and diarization pipelines, and it tops the Open ASR leaderboard among open-weight systems.

How to try it

Nemotron 3.5 Transcribe is an open-weight (OpenMDW-1.1) speech model you can run yourself for transcription, translation and speaker-labelled transcripts of recorded audio. NVIDIA's reference setup is a single H100 80GB.

  1. Set up the NeMo Speech environment as described on the model card.
  2. Download the checkpoint from Hugging Face (nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B).
  3. Load it with NeMo's ASRModel.from_pretrained and call transcribe() on a WAV file.
  4. For serving, follow the model card's vllm serve command and send audio through the OpenAI-compatible chat API.
python
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained("nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B")
transcriptions = asr_model.transcribe(["file.wav"])
Open the docs
Big newsRead on NVIDIA (Hugging Face blog)