Inkling AI: Top US Open-Weights Model, Specs & Cost

Inkling, a new AI model, scores 41 on the AA Intelligence Index, the highest for any US open-weights model. It uses a sparse Mixture-of-Experts design with 975 billion parameters, but only about 41 billion are active per inference. This allows it to handle context lengths up to 1 million tokens and process text, image, and audio inputs. What's interesting is Inkling trails some other models on reasoning benchmarks, but it uses far fewer tokens per task. Running a local version needs at least 290 gigabytes of RAM and keeps about 74% accuracy. The model costs $1.87 per million input tokens, which is about four times the median for open-weights models. The company behind Inkling, Thinking Machines Lab, is positioning it as a customizable foundation for enterprises and developers. They can fine-tune it through the Tinker training platform and own the resulting checkpoint. This approach is different from typical launches that focus on benchmark supremacy. The bottom line is Inkling offers a different path for AI development, focusing on customizability and efficiency rather than just top-tier benchmarks.