Meta brings Muse Wonder-trained models and spaces to Muse, Ai2 opens ToyBlocks, Ripple-65B ships open weights

Meta has replaced the models behind the Muse app and Meta AI. The new ones are specialized versions of its existing models, retrained with help from Muse Wonder, the flagship it launched in October. They inherit two Muse Wonder features: Active Memory, which keeps track of what still matters in a long conversation, and Active Focus, which looks for another route when a task gets stuck. Muse also got spaces the same day. You describe something like a gift tracker or a trivia game, Muse builds it, and it keeps itself updated from the apps you connect.

Meta says the share of tasks that correctly use critical context rose from 76% to 90%Vendor claim, and its own Personal Superintelligence Index went from 38% to 47%Vendor claim. Instead of keeping those abilities in the flagship alone, Meta used it to retrain the models already running inside its consumer apps. Teams building agents can borrow the way it measures memory and recovery. Spaces fit a wider trend, alongside Claude artifacts and ChatGPT apps, where the output is a small app that sticks around and keeps itself current instead of a chat reply.

ToyBlocks is Ai2's new simulation environment for training and testing robots. The world is deliberately small: a tabletop, some blocks and one robot arm. It generates pushing, flipping and stacking tasks on its own and produces demonstrations at scale, with no human teleoperating the arm. Skills are learned with reinforcement learning and motion planning in Isaac Lab, NVIDIA's open-source robot-learning framework, and a task-and-motion planner (TAMP) chains them into longer tasks. Code, data and models are all released.

Robot demos can look impressive while hiding why they fail. A narrow world lets researchers change the training data, rerun the same experiment and trace a failure to its cause. A team could, for example, add more pushing demos and see how the robot does on tasks it never trained on. Ai2 calls these controlled environments FoundationWorlds, and ToyBlocks is the first. For anyone working on robot policies, it is a cheap test bench for measuring how data volume and mix affect results.

Ripple-65B Preview is a 65B-parameter dense transformer (no mixture of experts) that Ricursive Intelligence, a company building AI for chip design, trained from scratch. It was pretrained on 12 trillion tokens and accepts up to 1 million tokens of context. The instruction-tuned model (BF16 and NVFP4) and a BF16 base model are on Hugging Face under Apache 2.0.

Ricursive reports 80.52Vendor claim on TerminalBench 2.1, ahead of US open-weight models more than ten times its size. That comparison covers US models only, though, and leaves out Chinese models like Kimi and DeepSeek. Because it's a 65B dense model, the company says the NVFP4 checkpoint fits on personal hardware like a DGX Spark. That gives teams one more commercially licensed option for running a coding agent in-house.