Tencent's Free Hy3 Rivals Flagship AI, Plus Claude's Hidden Mind & MIRA's World Model

Tencent has released Hy3, a 295 billion parameter Mixture-of-Experts model that's challenging proprietary models many times its size. Only 21 billion parameters activate per token, letting it support a 256,000 token context window without the compute costs of a genuine trillion-parameter system. Tencent says hallucination rates fell from 12.5% to 5.4% since the April preview, with commonsense errors nearly halved, after feedback from fifty internal product teams. The model is released under Apache 2.0, with weights on Hugging Face and a free API window on OpenRouter for two weeks. It's already deployed inside WeChat's assistants and Tencent's own game studios, another sign that open models are closing the gap on the budget-option label. Anthropic has published research suggesting Claude has something resembling a global workspace, the architecture cognitive scientists use to explain human conscious thought. Using a new interpretability tool called J-lens, researchers found that only a small portion of the model's internal activity is ever accessible to itself, echoing how little of our own brain activity we can consciously describe or reason about. Anthropic stresses this isn't a claim about subjective experience, but a functional similarity that matters for safety and interpretability research. The finding doesn't settle any philosophical questions, but it does suggest the internal structure of large models may share more with biological cognition than expected. General Intuition and Kyutai have built MIRA, a world model capable of rendering a full four-player Rocket League match in real time at twenty frames a second, without any physics engine or graphics code running underneath. The model was trained entirely on ten thousand hours of bots playing against each other, with no human match data involved, yet it keeps the ball, boost meters and each player's view consistent for minutes at a time. The code, training data and a playable demo have all been released openly. It's a striking demonstration that a model can learn the implicit rules of a physical system purely by watching it, without ever being told how it works. Google's Gemini Spark has gained the ability to monitor a topic continuously and respond the moment something changes, rather than waiting to be asked. Users can set it to track a stock price or a sports match, and Spark will send a full analysis as soon as a threshold is hit or an event ends. The feature is built into the desktop app that launched on Mac this month, alongside new third-party integrations and support for the Model Context Protocol. The idea is straightforward: instead of refreshing a page manually, let Spark do it automatically and flag what matters.