Inkling: A New Open-Weight 975B Moe with a Few Surprises (sebastianraschka.com)

Thinking Machines Lab has unveiled Inkling, a new open-weight 975 billion-parameter model that stands out in the competitive landscape of large language models (LLMs). While Inkling excels in specific evaluations such as IFBench and SimpleQA Verified, outperforming GLM-5.2 in these benchmarks, it shows limitations in reasoning and coding tasks. This nuanced performance indicates that Inkling is designed as a versatile all-rounder rather than specializing in any particular area. The architecture incorporates a sparse Mixture-of-Experts (MoE) design with conventional Transformer decoders, making it a notable entry in the landscape of large-MoE models. Significantly, Inkling introduces intriguing architectural features, such as the inclusion of small convolution layers that enhance local token mixing and a separate RMSNorm layer for improved token representations. Its unique approach to attention mechanisms—utilizing a sliding window with learned relative-position biases—may contribute to its adaptability across various tasks. Overall, though it may not dominate every benchmark, Inkling's broad performance profile and open availability for fine-tuning represent a valuable asset for the AI/ML community, encouraging further exploration and customization.