Inkling is a multimodal Mixture-of-Experts model from Thinking Machines Lab, built around a 975-billion-parameter design where only 41 billion parameters activate per token. Its architecture pairs a 66-layer decoder with hybrid local and global attention, and each sparse feed-forward layer routes a token to 6 of 256 experts alongside 2 shared experts. Inputs are processed through modality-specific encoders: a hierarchical patch encoder handles images and video, while audio is tokenized via discrete dMel units, then fused for text generation. Checkpoints are stored in BF16 precision and are published as fully open weights under an Apache license with a separate Model Acceptable Use Policy.

The model is designed for teams that need broad multimodal understanding combined with the flexibility of self-hosting or fine-tuning. NeMo AutoModel retains the checkpoint's fused expert layout and supports pipeline parallelism and expert parallelism for full-parameter fine-tuning, while a managed LoRA path through Tinker offers lighter adaptation work. Developers should note that the advertised one-million-token context window shrinks to 64K or 256K inside Tinker's managed LoRA fine-tuning path, which matters when planning long-context workflows. The open-weights release, multimodal breadth, and integration with established fine-tuning frameworks make Inkling a practical fit for research labs and production teams that want to customize a large mixture-of-experts model without rebuilding an inference stack from scratch.