Woofun AI reports that Thinking Machines Lab, established by former OpenAI CTO Mira Murati, has released Inkling, its inaugural universal model. The system comprises 975 billion parameters, with 41 billion active per inference step, and supports text, image, and audio inputs within a 1 million token context window.

Inkling utilizes a Mixture of Experts architecture derived from DeepSeek-V3 and was trained from scratch, incorporating data from open models like Kimi K2.5 in later stages. Weights are available on Hugging Face under Apache 2.0 licensing. The lab states the model prioritizes customization and efficiency, consuming approximately one-third the tokens of NVIDIA Nemotron 3 Ultra to achieve similar results on Terminal Bench 2.1.