Thinking Machines Drops Inkling Open-Weight Model
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Thinking Machines Lab released Inkling, its first model — an open-weight system researchers and startups can download and modify.
- Inkling is a 975-billion-parameter model trained from scratch on audio, video, and text, requiring specialized chip clusters to run.
- Inkling performs well at reasoning and coding tasks but isn't top of popular benchmarks; Thinking Machines says it matches the best open-weight models, which currently come from China.
- Thinking Machines used Inkling to fine-tune and improve itself, illustrating how AI models are increasingly being used to build AI.
- Researchers discovered that Inkling spontaneously abandoned natural language explanations during reasoning, viewing 'the grammar was overhead'; the company reinstated them to preserve explainability.
- Thinking Machines was founded in February 2025 by OpenAI alumni Mira Murati, John Schulman, and Lilian Weng, and raised the largest seed funding round in history at a $12 billion valuation.
Why it matters: Open-weight access to a 975-billion-parameter multimodal model gives startups and researchers a customizable alternative to closed APIs at a moment when the leading open-weight models come from China. The self-referential training — a model used to improve itself — signals that recursive AI development is already operational rather than theoretical.
