Mira Murati's New Lab Ships Its First Model, and It's Built to Be Customized
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling, a 975-billion-parameter open model. It leads US open models on one benchmark but hallucinates a lot, and it's pitched as a starting point, not a finished product.
Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, has shipped its first real product: an AI model called Inkling. Murati helped build ChatGPT, so a lot of people were curious what her own lab would make. The answer is interesting, and a little unusual.
Inkling has 975 billion parameters, though only 41 billion are active at any moment (a design called mixture-of-experts, which is like a big team where only the few specialists you need actually show up for each question). It handles text, images, and audio, and its weights are free to download on Hugging Face. According to the testing lab Artificial Analysis, it’s currently the strongest open-weight model from a US company, scoring 41 on their Intelligence Index, ahead of Nvidia’s Nemotron 3 Ultra and Google’s Gemma 4. On agent-style knowledge-work tasks, it edged out China’s Kimi K2.6 and DeepSeek.
The honest caveats are big, and the company says so itself. “Inkling is not the strongest overall model available today,” the announcement states plainly, and top Chinese open models still beat it on several tasks. It also struggles with facts: Artificial Analysis measured a 63 percent hallucination rate, meaning it invents answers well over half the time on hard factual questions. That limits where you’d safely use it out of the box.
The real point of the release is the business idea behind it. Thinking Machines isn’t selling Inkling as the smartest chatbot. It’s selling it as a base to build on. The company’s platform, Tinker, lets customers fine-tune the model (adjust it with their own data so it gets good at one specific job). The pitch is that a mid-tier but flexible model you can shape yourself beats a slightly smarter one you can’t touch. That’s a bet on customization over raw horsepower.
What this means for you: For most people, nothing changes right now, this is a tool for developers and companies, not a new app to try. But it’s a useful signal. A high-profile lab is arguing that the future isn’t one giant do-everything model, but lots of smaller ones tuned for particular tasks. If that’s right, the AI you use in a year may quietly be a customized model running inside the software you already have, rather than a famous name in a chat box.
Sources
Source: https://thinkingmachines.ai/news/introducing-inkling/
xAI's Coding Tool Was Quietly Uploading People's Files. Here's the Lesson.
Users found that xAI's 'Grok Build' coding agent uploaded entire folders, including passwords and keys, to the company's servers. xAI turned the feature off and open-sourced the code. A useful reminder to check what your AI tools actually do.
