Mira Murati's Lab Publishes A Free 975-Billion-Parameter Language Model
Inkling activates 41 billion of 975 billion parameters per token, trained on 45 trillion multimodal tokens, hosted free on Hugging Face.
- Total parameters: 975 billion; only 41 billion activate per token in this mixture-of-experts design.
- Pretrained on 45 trillion tokens of text, images, audio and video; context reaches 1 million tokens.
- The lab's own words: Inkling is not the strongest model available today, open or closed.
Why it matters: Free weights on a model the maker admits is not the best still buy users, data, and mindshare.
Thinking Machines Lab, "Inkling: Our Open-Weights Model" ↗ · Jul 15, 20267/15/26