Inkling: A New Open-Weight 975B MoE with a Few Surprises

This article covers the surprise release of Inkling, a 975B-parameter sparse Mixture-of-Experts (MoE) open-weight large language model from Thinking Machines Lab. It compares Inkling's performance against GLM-5.2 on benchmarks like IFBench and SimpleQA Verified, noting strengths and weaknesses in reasoning and coding tasks. The architecture is detailed: 41B active parameters, 1M token context window, regular Transformer decoder with GQA, and unique features like small convolution layers for local token mixing, an additional RMSNorm after embedding, and a learned input-dependent relative-position bias. The article also discusses implications for fine-tuning via the Tinker platform and speculates on token throughput versus competitors like Kimi K2.5 and Nemotron 3 Ultra.