Mira Murati Unveils First AI Model Since Leaving OpenAI

For her first technological speech since leaving OpenAI in September 2024, Mira Murati makes a big impact. The former chief technical officer has just unveiled the first AI model from her startup Thinking Machines Lab with a fully assumed ambition: to bet on open source to redefine corporate data sovereignty. In brief Thinking Machines publishes Inkling, its first artificial intelligence model. Mira Murati, former CTO of OpenAI, leads this new AI-focused lab. Inkling takes an open source approach to encourage contributions from researchers and developers. The initiative stands out from the strategies favored by several major AI players. A new AI breaking with the closed models paradigm Named Inkling, Mira Murati’s AI model is based on a Mixture-of-Experts (MoE) architecture. According to the official Thinking Machines announcement, it has 975 billion parameters. However, it only activates 41 billion per token of processed text. The result: a artificial intelligence that is more performant, faster, and less costly to run than a traditional massive model. Designed as a customizable base via the Tinker tuning tool, Inkling natively accepts text, visual, and audio data. It handles context that can reach 1 million tokens, about 750,000 words. This revolutionary AI model was even pre-trained on 45 trillion tokens. But its greatest particularity lies in its mode of distribution. The full weights are indeed published for free on Hugging Face under an Apache 2.0 license, without restrictions on commercial use. Your 1st cryptos with BitgetThis link uses an affiliate program. Decoding: any company can download, host, and modify locally the AI algorithm. A software decentralization approach that breaks with the technical and economic dependence on proprietary APIs from OpenAI and Google! In this context, Thinking Machines highlights an important point in its official release: Inkling is primarily aimed at research and experimentation. The AI model is therefore at the community’s disposal. In other words, researchers, developers, and engineers can study it, improve it, and design new applications. An approach that fully reconnects with open source principles. With Inkling, Mira Murati reignites the AI infrastructure war On MCP Atlas, a benchmark that measures the reliability of an AI agent to perform real tasks via the standard MCP (Model Context Protocol), the AI model scores 74.1%, compared to 44.7% for Nvidia’s Nemotron 3 Ultra. This is its main Western competitor in open source AI. On SWE-Bench Verified, which evaluates an agent’s ability to autonomously fix real software bugs on GitHub, it reaches 77.6%. It also takes the lead over Nemotron. On Teminal Terminal Bench 2.1, Inkling is largely outperformed by Z.ai lab’s GLM 5.2: 63.8% versus 82.7%. Kimi K2.6 also dominates Humanity’s Last Exam, a scientific reasoning test at a doctorate level.