Inkling is a multimodal Mixture-of-Experts model from Thinking Machines Lab, built around a 66-layer decoder that blends hybrid local and global attention with sparsely activated feed-forward blocks. Each token is routed to six of 256 experts while two shared experts are always evaluated, yielding a 975B-parameter architecture with roughly 41B parameters active per token. Inputs are handled through modality-specific encoders: text is processed natively, images and video pass through a hierarchical patch encoder, and audio is decoded via discrete dMel tokens, with the fused expert layout preserved for downstream fine-tuning. The checkpoint is published in BF16 precision on the Hugging Face Hub under the thinkingmachines organization, and NeMo AutoModel documents pipeline and expert parallelism recipes for adapting it to medical visual question answering datasets such as MedPix-VQA.

For practitioners, Inkling is positioned as an open-weight foundation that combines broad modality coverage with the efficiency benefits of sparse activation, making it attractive for teams that want to fine-tune a very large model without paying full dense inference costs. The hybrid attention pattern and shared-expert routing are well suited to long-context reasoning across mixed text, visual, and audio inputs, while the released recipes lower the barrier for multi-node training and specialization. Teams evaluating general-purpose multimodal assistants, document or video understanding systems, or domain-specific adaptations such as clinical imaging stand to benefit from a checkpoint that exposes both the full architecture and reproducible fine-tuning pipelines, with the open-weights status enabling self-hosted deployment and further research.