Inkling: Thinking Machines Lab Releases 975B Open-Weights Model

Thinking Machines Lab has released Inkling — an open multimodal model with 975 billion parameters that understands text, images, and audio and supports a context window of up to 1 million tokens.

The MoE (mixture-of-experts) architecture keeps it practical: only 41 billion parameters are active per request, so the model doesn't require a supercomputer. Weights are already posted to Hugging Face for download and fine-tuning on specific tasks.

An open model at this scale disrupts the usual hierarchy: everything that previously required closed, paid APIs can now be run independently or fine-tuned on proprietary data. The race between closed and open models is heating up with every passing week.