Mira Murati’s startup, Thinking Machines, targets enterprise customisation over benchmark leadership with Inkling, its first multimodal open-weights model.
On 15 July 2026, Thinking Machines Lab, a San Francisco-based AI startup founded by Mira Murati, former CTO of OpenAI, launched its first open-weights AI model, Inkling. Released as an open-weights model, it allows developers and enterprise teams to download, inspect, host, and fine-tune the system locally on their own infrastructure.
The model is built as a Mixture-of-Experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters per query. It supports up to a 1-million-token context window and was trained from scratch on 45 trillion tokens encompassing a multimodal blend of text, images, audio, and video. Inkling processes text, vision, and audio natively without relying on separately bolted-on vision adapters.
Users can explicitly adjust or dial the model’s reasoning and thinking effort up or down depending on whether they prioritise speed and cost or deeper chain-of-thought accuracy. It is trained to deliver calibrated responses and withstand censorship, exhibiting non-compliance with external censorship requests on evaluated benchmark datasets. Inkling is available on Thinking Machines’ Tinker customisation console, designed for post-training and fine-tuning workloads, and was released alongside a preview of Inkling-Small, a lighter variant featuring 12 billion active parameters.
Thinking Machines explicitly acknowledged that Inkling is not aimed at beating the single highest benchmark score against closed models such as GPT-5 or Claude. Instead, it is designed to serve as a broad, well-rounded base for customisation, offering Western developers and enterprise teams a high-capability US-built open-weights foundation alongside leading Chinese open models like Qwen.
