Thinking Machines Lab Releases Open-Source MoE Model Inkling: 975B Total, 41B Active

Thinking Machines Lab launched Inkling, its first fully open-weight foundational model family. Using a Mixture-of-Experts architecture with 975B total parameters and only 41B active per token, it reduces inference costs significantly. The model handles text, image, and audio inputs, outputs text, and supports a context window up to 1M tokens. It is open-sourced under Apache 2.0. Inkling is available on platforms like Tinker and Hugging Face, with support from vLLM, SGLang, Modal, Baseten, and Databricks. Notable figures like Mira Murati, Soumith Chintala, John Schulman, and Lilian Weng highlighted its open weights, customizability, and practical focus. Independent reviews call it the strongest open-weight model from the US, though it lags behind top Chinese open models and the best closed-source models on some benchmarks. For developers and enterprises needing to deploy or fine-tune large models, Inkling provides a high-performance, freely customizable multimodal option.