Meta AI Updates: October 22, 2026

1. TorchTitan Adds TitanRL, Muse Spark Open Support, and Faster Checkpointing

Meta. The PyTorch TorchTitan team released a broad update to its PyTorch-native training platform, headlined by TitanRL, which reuses one model definition from pretraining through post-training and inside vLLM, runs asynchronous rollouts, and offers a batch-invariant mode in which trainer and generator produce bitwise-identical log probabilities. New model support covers Muse Spark Open, Muse Glimmer, Qwen3-VL, Qwen3.5, Kimi K2.7, Kimi K3, and DeepSeek-V4 across its 5D parallelism stack. The team reports about 5,300 tokens/s per GPU (roughly 1,550 TFLOPs/GPU) training DeepSeek V3 671B in MXFP8 on 256 GB300s. Its new FlexShard library cuts Muon optimizer time by up to 79% on 256 H100s, and the new torch_checkpointing library drops checkpoint write time on DeepSeek V3 at 512 H100s from 384 s to 13 s. Source

2. Meta Says Muse Has Over 10 Million Users, 25% of Them Daily

Meta. On its Q3 earnings call, Meta said Muse has more than 10 million users and about 25% of those who set it up use it daily, with shopping and email/calendar productivity as top uses. Meta plans to keep Muse free for most users, take a cut of commerce, and charge power users for advanced models or more tokens. It expects to cut Muse’s serving cost by more than 10x within a year and previewed a “Muse Confidential VM” designed to keep even Meta from seeing user content. Source