Mira Murati Left Her First AI Model After Leaving OpenAI—And It’s Open Source

Mira Murati left OpenAI in September 2024 to pursue her own interests. About two years later, that investigation was sent. Thinking Machines Lab, the company he founded, was spun off Inkling-a multimodal AI model trained from scratch, with any weight available for free download.

When the OpenAI team removed Sam Altman in November 2023Murati, then CTO, was appointed interim CEO. Altman was brought back five days later, Murati returned to CTO, then left about 10 months after that. He founded the Think Machines Lab in February 2025.

Then the company became quiet—and rich. They raised $ 2 billion at a cost of $ 12 billion in July 2025, led by Andreessen Horowitz and Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street on the side – one of the largest seeds in the history of Silicon Valley. at that time.

Reports in November 2025 had the company seeking a new round at a cost of $50 billion. Those stories he has fallen by January 2026.

What is Inkling?

Inkling is a hybrid expert-architecture model where only the network component handles any given task, keeping ideas fast without losing depth. It’s a very large model: It has 975 billion parts (the internal parameters that define how the model processes information), and 41 billion active parts, so forget about running it on your machine.

Being multimodal, the model accepts texts, images, and words, and supports a window of information – the amount of text that this model can consider at the same time – 1 million tokens, about 750,000 words. It’s also fixed at 45 trillion tokens based on text, images, audio, and video.

“Our first model, Inkling. Trained from the beginning, rich and open, well developed on Tinker today,” Murati wrote on X. The fact that it is taught from the beginning means a lot, especially in the open world because it can bring fresh air to Western developers who care about China but have to use Asian models for their developments because the top AI companies in the West are focused on sending local models.

Optimization is the process of retraining an existing model on a specific set of data to improve its performance for a specific task. Tinker is a Thinking Machines cloud platform built around that application. All the weights are there too Hugging Face under the Apache 2.0 license, there are no restrictions.

Inkling’s clearest successes come in practical applications. On the MCP Atlas—which measures how an AI agent completes real-world tasks using the Model Context Protocol, an open standard for connecting AI agents to external tools and services, given as the number of tasks completed—Inkling records 74.1%. That’s about 30 points higher than Nvidia’s Nemotron 3 Ultra, the West’s biggest open-source competitor.

On SWE-Bench Verified—testing whether an AI agent can randomly fix real GitHub bugs, expressed as the number of issues resolved—Inkling scores 77.6%, well above Nemotoron’s 70.7%.

Overall, Thinking Machines is marketing this model as “well-rounded” and casual. It means that it does not interfere with quality in one type of work because its ability focuses on something else (such as examples that are very good at writing but absorb writing, for example).

Chinese brands still have limitations in several areas. Z.ai’s GLM 5.2 they scored 82.7% on Terminal Bench 2.1—a benchmark for measuring autonomous AIs in real-world environments, scoring the number of tasks completed—versus Inkling’s 63.8%. Kimi K2.6 presides over Humanity’s Last Exam, a test of PhD scientific ideas.

The Thinking Machine accepts this. Inkling is not the most powerful brand available today, open or closed.

In fact, it is the best quality test made by a Western lab. Developers who – for legal, security, or compliance reasons – cannot run large projects through Beijing-built models now have an alternative. Chinese auto parts.

Now, these manufacturers have a model that (although worse than the best Chinese models in almost every way) fits well with their ideas, expectations and values. Sequential files may make this model more successful in certain applications, making the files more competitive in benchmarks against Asian models.

On the FORTRESS Adversarial—which measures how well an analogy consistently rejects harmful information without blocking valid, scoring as a percentage of the correctness—Inkling scores 78.0%, the highest mark among all types of weights in the comparison.

Next to Inkling, Thinking Machines predicted Inkling-Small: 276 billion total units, 12 billion active, already comparing with the big version in many terms. The weights arrive after the test is completed, with no time limit.