Overview
- Thinking Machines publicly released Inkling on Wednesday, July 15, and made full model weights available on Hugging Face and partner platforms.
- Inkling is a mixture-of-experts system with 975 billion total parameters and about 41 billion active per token to cut per-query compute.
- The model is multimodal, trained on a large corpus across text, image, audio and video, and was trained on Nvidia GB300 (Vera Rubin) systems.
- Inkling is positioned as an enterprise base for customization through the Tinker platform and distribution on Databricks and other hosts rather than a metered API product.
- Public benchmarks show Inkling is a strong generalist that excels at agentic and coding tasks but does not top the best open or closed models, and the company used other open models to bootstrap some early post-training data.