Mira Murati's new 975B model Inkling isn't even trying to beat OpenAI

Former OpenAI CTO Mira Murati launched her startup's inaugural base model, Inkling, built on a 975-billion parameter Mixture-of-Experts architecture with 41 billion active parameters per token. Trained on a massive dataset of 45 trillion multimodal tokens including text, audio, image, and video, the model supports a context window of up to 1 million tokens.

Alongside the flagship release, Thinking Machines Lab introduced Inkling-Small, a lightweight variant featuring 12 billion active parameters designed for cheaper and faster local inference. While tech startups usually spend millions pretending their new algorithm is godlike, the team openly admitted that Inkling does not hold the top spot on any major benchmark, positioning it instead as an adaptable open-weights foundation.

The core selling point rests on Tinker, an automated infrastructure platform designed to handle dataset uploading, fine-tuning, evaluation, and weight deployment without requiring engineers to babysit machine learning clusters. To demonstrate this pipeline, Inkling was instructed to generate its own fine-tuning task—specifically training itself to never use the letter 'e'—before automatically evaluating the output and switching to the modified weights.

By offering an alternative to expensive closed APIs from Anthropic and raw open-source weights from foreign labs, the startup aims to establish automated custom model factories for corporate clients.

The era of worshipping synthetic benchmark leaderboards might finally be giving way to useful, ugly corporate automation. Time will tell if enterprises actually want a self-tuning AI factory or if they just want a cheaper chatbot that doesn't hallucinate fake spreadsheets.