Thinking Machines Lab released Inkling, a 975B-parameter sparse Mixture-of-Experts open-weight LLM with 41B active parameters and a 1M-token context window. Benchmark comparisons with GLM-5.2 show Inkling leads on IFBench and SimpleQA but trails on reasoning and coding-agent tasks. Architecturally, Inkling stands out for three design choices: short kernel-4 convolutions after key/value projections and branch outputs for local token mixing; an additional RMSNorm directly after the embedding layer; and a learned input-dependent relative-position bias instead of RoPE, which Thinking Machines claims extrapolates better to longer sequences. Of 66 decoder layers, 55 use local attention with a 512-token window, and the 11 global layers apply the positional bias only over the preceding 1,024 tokens. The model is positioned as a base for fine-tuning rather than a benchmark leader, aligning with Thinking Machines' Tinker customization platform.
