Thinking Machines Lab drops Inkling: open-weight model that self-improves
975B parameter model trained to fine-tune itself, dropping grammatical overhead in reasoning.
Thinking Machines Lab, the AI startup founded by former OpenAI leaders including Mira Murati, John Schulman, and Lilian Weng, has released its first model, Inkling. The open-weight model boasts 975 billion parameters and was trained from scratch on audio, video, and text inputs. While it doesn't top popular benchmarks, the company claims Inkling performs well on reasoning and coding tasks. Notably, the lab used Inkling to fine-tune itself—a sign of AI's growing role in its own development. During training, the model's chain-of-thought reasoning became more concise over time, dropping grammatical overhead while remaining comprehensible. This self-improvement could lower deployment costs and make the model more efficient.
Inkling positions Thinking Machines as a legitimate player in the open-weight AI race, challenging Chinese models that currently dominate the category. The startup received a record $12B seed valuation and has already released a fine-tuning tool (Tinker) and voice interaction capabilities. The open-weight release aligns with the company's vision of decentralizing AI control, contrasting with closed models from OpenAI and Anthropic. Anthropic recently filed for an IPO that could value it over $1T. By offering a performant, modifiable model, Thinking Machines aims to give researchers and startups an alternative to proprietary systems, potentially reshaping the competitive landscape.
- Inkling is a 975B parameter open-weight model trained on audio, video, and text, rivaling leading Chinese open-weight models.
- The model was used to fine-tune itself, resulting in chain-of-thought reasoning that became more concise and efficient over time.
- Company founded by ex-OpenAI CTO Mira Murati, co-founder John Schulman, and VP Lilian Weng, with a $12B seed valuation.
Why It Matters
Open-weight Inkling democratizes AI development, challenges Chinese dominance, and showcases self-improvement that could reduce costs.