Inkling is part of the ling family of multimodal mixture-of-experts models released by Thinking Machines Lab with open weights, designed as a general-purpose system for developer-facing applications such as coding assistants, agentic and tool-use workflows, chatbots, and retrieval-augmented generation. The full variant carries 975 billion total parameters with 41 billion active per token, while Inkling-Small scales this design down to 276 billion total parameters with 12 billion active, giving teams a choice between deeper reasoning capacity and lighter deployment. Both checkpoints are published on Hugging Face under an Apache 2.0 license, making them suitable for fine-tuning and integration into third-party products.
The model architecture is a multimodal autoregressive transformer that combines a 66-layer decoder with hybrid local and global attention, using a 512-token sliding window and global attention every sixth layer, eight key-value heads, and short convolutions with a kernel size of four. Each sparse feed-forward block routes tokens to six of 256 experts alongside two shared experts, which is what enables the large total parameter count while keeping active compute moderate. Inputs flow through a hierarchical patch encoder for images and video and discrete dMel tokens for audio, so the same backbone handles text, vision, and sound without separate specialist paths, and NeMo AutoModel retains the fused expert layout for distributed training and inference.
