Milestones
- Inkling released as open-weight15 Jul 2026Complete.
- Inkling-Small variant releasedJul 2026Complete.
Demonstrate that frontier-quality multimodal models can be trained and released openly, empowering independent researchers and builders to customize and extend frontier AI.
First release deployed. Open-weight multimodal mixture-of-experts model trained from scratch.
Updated 15 Jul 2026Checked 11 Oct0 updates this week
Expert sub-networks handle different reasoning; the model routes tokens selectively, activating only 41B of 975B parameters per token.
A single architecture processes text, images and other inputs without separate encoders.
Released under a permissive license, enabling researchers and teams to fine-tune and deploy it.
| Spec | Inkling |
|---|---|
| Total parameters | 975BR (reported) |
| Active parameters per token | 41BR (reported) |
| Architecture | Multimodal mixture-of-expertsR (reported) |
R reported by the company
Thinking Machines Lab
Tinker fine-tuning service, AI research
Thinking Machines Lab was founded in February 2025 by Mira Murati, OpenAI's former CTO, with researchers from OpenAI, Meta and Mistral. It publishes research and offers Tinker, a service for customising open-weight models.
thinkingmachines.ai (opens thinkingmachines.ai)Company profile →