Clones voices from audio samples and generates steerable speech in 43 languages. Its sparse MoE activates about 100 million parameters per token, enabling real-time use in browsers and on local devices.
Runs a 145B agent model on a single consumer GPU using sparse latent MoE. The NVFP4 build reports 98.3 tok/s agent decoding on an RTX 5090 D, while also relying on host RAM.
Speeds up lossless speculative decoding for vision-language models by focusing drafter training on visually dependent tokens. Sparse attention also reduces redundant work on long video inputs; reported speedups reach 3.58×.