Inkling: Our open-weights model
Summary
Inkling is presented as an open-weights Mixture-of-Experts transformer with 975B total parameters (41B active) and a context window of up to 1M tokens. The article highlights its multimodal capabilities, self-finetuning demonstrations, benchmarking, and availability on the Tinker platform and Hugging Face. It emphasizes customization, safety, and scalable RL-based improvements for open-weight models.