Thinking Machines Unveils Inkling: An Open Weights Multimodal AI Designed for Strategic Specialization

Thinking Machines, the AI startup founded by former OpenAI CTO Mera Miati and backed by a $2 billion investment from A16Z, has officially launched Inkling, a fully open-weights multimodal model trained from scratch. The company, which garnered a $12 billion valuation prior to shipping a product and brought in key talent like OpenAI co-founder John Schulman and VP of Research Barrett Zoff, previously introduced Tinker, an API for fine-tuning open-weight models. Inkling, a mixture of experts (MoE) model, features 970 billion total parameters, dynamically activating 41 billion per token for efficient compute. It was pre-trained on an extensive 45 trillion tokens of text, images, and audio, handles a 1 million token context window, and is Apache licensed with weights readily available on Hugging Face.

Despite its advanced architecture, Inkling deliberately does not aim to outperform frontier models like Fable 5 or GPT 5.6 Sol on raw benchmarks, openly admitting to landing mid-table even among open Chinese models. This strategic positioning is by design, with Thinking Machines emphasizing Inkling’s unique capabilities for specific use cases rather than raw intelligence. Key features include a “thinking effort” dial that allows users to adjust between cheap, instant answers and highly accurate results, matching Neotron 3 Ultra on benchmarks while using a third of the tokens. Inkling also processes raw audio and pixels directly, bypassing traditional encoder models, and incorporates “epistemics,” rewarding the model for admitting uncertainty, making it a leading model for forecasting future events, surpassing GPT 5.5 and Opus 4.8. The company’s vision is to provide Inkling as a capable, freely available base model, driving revenue through specialized fine-tuning services via Tinker, transforming it into highly effective specialist agents for targeted problems.