Nearly two years after she stepped away from OpenAI, Mira Murati has finally shipped the thing she had been quietly building. Her company, Thinking Machines Lab, has launched its first model, Inkling, a multimodal AI system trained entirely from scratch, with every weight available for free download. That last detail is what makes the release stand out, because the biggest AI companies in the Western world have mostly focused on shipping closed-source models.
"Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today," Murati wrote on X. The fact that it was trained from scratch carries real weight. In the open-source community it could be a breath of fresh air for Western developers who are wary of China but have had to lean on Asian models for their own work.
The exit that led here
When OpenAI's board fired Sam Altman in November 2023, Murati, then the CTO, was named interim CEO. Altman was reinstated five days later and Murati returned to the CTO role. Roughly 10 months after that, in September 2024, she left OpenAI for good to pursue her own path. She founded Thinking Machines Lab in February 2025.
Quiet, and very well funded
The company then went silent, but it also grew rich in the meantime. In July 2025 it raised $2 billion at a $12 billion valuation. The round was led by Andreessen Horowitz, with Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street alongside. At the time it was one of the largest seed rounds in Silicon Valley history. In November 2025 word emerged that the company was chasing a new round at a $50 billion valuation, but those talks collapsed by January 2026.
Inside how Inkling is built
Inkling is a mixture-of-experts model, an architecture in which only a portion of the network activates for any given input, keeping inference fast without sacrificing depth. It is a very big model. It carries 975 billion total parameters, the internal settings that define how the model processes information, with 41 billion active per task, so forget about running it on your local machine.
Because it is multimodal, Inkling accepts text, images, and audio. It supports a context window of 1 million tokens, the amount of text it can reason over at once, which works out to roughly 750,000 words. It was pretrained on 45 trillion tokens spanning text, images, audio, and video.
Open weights, and why they matter
Fine-tuning is the process of retraining an existing model on a specialized dataset to sharpen its performance on a specific task. Tinker is Thinking Machines' cloud platform built around exactly that use case. The full weights are also on Hugging Face under an Apache 2.0 license, with no restrictions attached.
Where Inkling pulls ahead
Inkling's clearest wins come in agentic tasks. On MCP Atlas, which measures how reliably an AI agent completes real-world tasks using Model Context Protocol, the open standard for connecting AI assistants to external tools and services, and is scored as a percentage of tasks completed, Inkling posts 74.1%. That is nearly 30 points above Nvidia's Nemotron 3 Ultra, the main Western open-weights rival in the comparison. On SWE-Bench Verified, a test of whether an AI agent can autonomously fix real GitHub software bugs, scored as a percentage of issues resolved, Inkling scores 77.6%, again above Nemotron's 70.7%. Thinking Machines is selling the model as "well-rounded" and generalist, meaning it does not sacrifice quality on one set of tasks just to shine on another.
Where China still leads
Chinese models still hold the edge on several fronts. Z.ai's GLM 5.2 scores 82.7% on Terminal Bench 2.1, a benchmark that measures autonomous AI coding agents in a real terminal environment as a percentage of tasks completed, against Inkling's 63.8%. Kimi K2.6 leads on Humanity's Last Exam, a test of PhD-level scientific reasoning. Thinking Machines acknowledges this openly. Inkling is not the strongest model available today, open or closed.
What it is, though, is the most capable open-weights model built by a Western lab. Developers who, for legal, security, or compliance reasons, will not route workloads through models built in Beijing now have a real alternative to self-hosting Chinese systems. Even if it trails the best Chinese models at almost everything, it aligns better with those developers' ideals, expectations, and values. Later fine-tunes can make it excel at specific tasks, making them competitive in benchmarks against Asian models.
Safety, and a smaller sibling
On FORTRESS Adversarial, which tests how consistently a model refuses genuinely harmful prompts without over-blocking legitimate ones, scored as a percentage correctly handled, Inkling scores 78.0%, the highest mark among all open-weights models in the comparison. Alongside Inkling, Thinking Machines previewed Inkling-Small, with 276 billion total parameters and 12 billion active, already matching the larger model on most reasoning benchmarks. Its weights will arrive once testing is complete, with no timeline given.
