Thinking Machines Lab Releases Inkling Open-Weights Multimodal Foundation Model

Thinking Machines Lab has released Inkling, its first open-weights foundation model. Designed as a customizable base model rather than a benchmark-focused flagship, Inkling is a multimodal Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters. It supports text, image, and audio inputs and is released under an Apache 2.0 license.

The model features several distinct architectural choices, including the use of relative positional encoding instead of the more common RoPE, as well as scaled-up short convolution layers. It was trained from scratch on approximately 45 trillion tokens. While it is not intended to be the most powerful model on every benchmark, it has received praise for its concise reasoning and strong tool-calling capabilities, making it a practical choice for developers looking to build and fine-tune their own agentic applications.

Inkling is available immediately on the company’s Tinker platform and Hugging Face, with significant day-zero support from major infrastructure providers like vLLM, Modal, and NVIDIA. Industry analysts view the release as a strategic move to provide a high-quality, U.S.-based open-weights alternative to the recent surge of powerful models from Chinese labs. By prioritizing a transparent, tunable foundation over raw leaderboard performance, Thinking Machines Lab aims to establish a reliable substrate for future research and commercial development.