People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way. We share our approach, early results, and a quick look at our model in action. thinkingmachines.ai/blog/int…

Big congrats to our SAM 3D team for receiving a Best Paper Honorable Mention at #CVPR26! This prestigious recognition underscores their incredible work pushing the boundaries of computer vision. Read the paper here: arxiv.org/abs/2511.16624

Cosmos 3 ties everything together. Previous releases separated world generation, physical understanding, and controlled scene generation. Cosmos 3’s MoT architecture unifies these capabilities by pairing an autoregressive reasoner tower with a diffusion-based generator tower.

🚨New work: Provable Scaling Laws of Feature Emergence from Learning Dynamics of Grokking (arxiv.org/abs/2509.21519) In this work we propose a mathematical framework, named Li2, that explains the dynamics of grokking (i.e., delayed generalization) in 2-layer nonlinear networks. Specifically, it 1️⃣ Tells exactly what features will emerge during training. 2️⃣ Gives provable scaling laws of generalization/memorization, i.e. O(M log M) data samples suffice for generalization behavior of group arithmetic task of order M group. 3️⃣ Provides a more fundamental explanation for the popular empirical hypothesis that "generalization circuits learn slower but is more efficient than memorization circuits". So how?

wrote a guide on getting compute grants as a student, something I wish I did more at the beginning of my PhD. It's honestly one of the highest ROI things you can do as a student (we've gotten 100k gpu hrs for roughly 2 weeks of work writing). nightingal3.github.io/blog/2…

ThreadWeaver Adaptive Threading for Efficient Parallel Reasoning in Language Models

We identified an issue with the Mamba-2 🐍 initialization in HuggingFace and FlashLinearAttention repository (dt_bias being incorrectly initialized). This bug is related to 2 main issues: 1. init being incorrect (torch.ones) if Mamba-2 layers are used in isolation without the Mamba2ForCausalLM model class (this has been already fixed: github.com/fla-org/flash-lin…). 2. Skipping initialization due to meta device init for DTensors with FSDP-2 (github.com/fla-org/flash-lin… will fix this issue upon merging). The difference is substantial. Mamba-2 seems to be quite sensitive to the initialization. Check out our experiments at the 7B MoE scale: wandb.ai/mayank31398/mamba-t… Special thanks to @kevinyli_, @bharatrunwal2, @HanGuo97, @tri_dao and @_albertgu 🙏 Also thanks to @SonglinYang4 for quickly helping in merging the PR.