If you’ve been following AI for even a few months, you’ve probably noticed a pattern. Every week there’s a new paid AI tool promising to do everything faster, better, and cheaper—right up until the subscription page loads. Meanwhile, quietly, in GitHub repos and research blogs, open-source models are improving at a pace most people completely miss.

UNAM runs the largest university in Mexico. Every year, hundreds of thousands of students take an entrance exam that determines whether they get in. This year, for the first time, the whole thing went remote. They deployed a lockdown browser, AI webcam monitoring, and one human supervisor per 150 applicants. The kind of setup that sounds serious on paper. Then the scores came in. Students hitting 100 or above jumped from 3.5 percent in previous years to 16.3 percent this year. At the very top end, scores of 110 or higher went from 0.9 percent to 5.5 percent. Not a small shift. Not noise. A roughly fivefold increase in top scores, in one year, under one new format. An expert commission investigated. Their conclusion: administer the entire exam again, in person, to around 58,000 people. The rector apologized to students who hadn't cheated. They now have to prepare for and sit another exam anyway.

Mark Zuckerberg published a 14-page essay today about why open-source AI is the path forward for humanity. Distribute intelligence rather than centralize it. Put the power in everyone's hands. A new era of personal empowerment. On the same day, Meta released Muse Glimmer, an open-source version of its most powerful model, Muse Spark, that anyone can download, modify, and build on for free. But the interesting part is, Muse Spark itself stays closed. You still pay to access it. The open version is nearly identical, Meta says, but the model that actually competes at the frontier, the one Zuckerberg's essay is implicitly defending remains behind a paywall. That gap between the philosophy and the product decision is what makes today's announcement interesting.

Nvidia just dropped a 120B model that only uses 12B parameters at a time. Take a second with that. You get the reasoning depth of a 120B model. You pay the compute cost of a 12B one. That gap is not a rounding error or a marketing trick. It is the whole point of what Nemotron 3 Super is built to do. This is not another chatbot release. Nvidia built this specifically for AI agents — systems that plan, call tools, check their own work, and run for hours without a human in the loop. The use case is different. The architecture is different. And if you are building anything with agents in 2026, the timing of this release is hard to ignore. It's already live. Weights are on HuggingFace. Let's get into what actually makes it interesting.

Just a few years ago, producing high-quality video, audio, or digital experiences required massive budgets and specialized teams. Now, that's changing dramatically. AI has evolved from a simple productivity tool to the core engine of creative production. We're witnessing a shift where artificial intelligence isn't just assisting creators, it's fundamentally reimagining how creative work is executed. Advanced AI models are now generating complex multimedia content that goes far beyond basic text-to-image tricks. Some of these models can run on local machines, others are open for experimentation, and many are redefining the boundaries of real-time, interactive media generation. If you're a creator, artist, or someone exploring how AI can fit into your creative workflow, understanding these models is crucial. They're the foundation of the next generation of creative tools which gives you the power to generate video, shape audio & experiment with digital experiences.

The mixture-of-experts trick changed how people think about LLMs. Instead of running every parameter on every token, you activate a small fraction of the network per forward pass and somehow the quality stays competitive while the compute drops. It's the reason models like Mixtral punched above their weight. Everyone in the LLM space understood it immediately. Nobody had done it openly for image generation. Until now. Nucleus-Image is a 17B parameter diffusion transformer that activates roughly 2B parameters per forward pass. It beats Imagen4 on OneIG-Bench, sits at number one on DPG-Bench overall, and matches Qwen-Image on GenEval. It's also a base model. No fine-tuning, reinforcement learning or human preference tuning. What you're seeing in those benchmarks is raw pre-training performance. That's either impressive or a caveat depending on what you need it for, probably both.

For deep research tasks, the options are mostly proprietary. Perplexity, ChatGPT DeepResearch, paid tools that do the job but keep your data on their servers and charge you monthly for the privilege. Yes you can use open source reasoning models like DeepSeek-R1 or Qwen3 for complex analysis and they are genuinely capable. But they are not built specifically for agentic deep research. They reason well. They do not orchestrate. That gap is exactly what MiroThinker 1.7 is designed to fill. An open source model built from the ground up for long horizon research tasks, step by step verification and up to 300 sequential tool calls without losing the plot. If you handle sensitive research and cannot pipe it through a third party server, this is worth paying close attention to.