Inkling

Sebastian Raschka's architecture analysis details that Inkling is a 975B-parameter open-weight Mixture-of-Experts model that activates 41B parameters per token and supports a context window of up to 1,048,576 tokens, placing it in the same general size class as Kimi K2.5 and GLM-5.2. The model has 66 decoder layers wit The architecture introduces several less-common details: 55 sliding-window attention layers and 11 global attention layers in a repeating 5-to-1 local-global pattern, with local layers using a 512-token window and 4-to-1 grouped-query attention while global layers use 8-to-1 GQA. Inkling skips RoPE in favor of a learne

Thinking Machines Lab has released Inkling, described as its first open-weights model and a mixture-of-experts (MoE) multimodal system that natively accepts image and audio inputs alongside text. The model is published under the permissive Apache 2.0 license, one of the least restrictive terms an AI lab can choose, mea The MoE architecture activates only a subset of parameters for any given input, a sparse routing approach that lets teams scale total capacity while keeping per-forward-pass compute cost manageable. Pairing that architecture with native image and audio handling places Inkling in the growing category of open models buil

Thinking Machines Lab has released Inkling, its first model, as an open-weight system trained from scratch to handle audio, video, and text inputs. WIRED reports the model has 975 billion parameters and is designed for advanced reasoning and coding, though it requires a cluster of specialized chips to run. The lab nota The release positions Thinking Machines Lab — founded by former OpenAI executives including Mira Murati, John Schulman, and Lilian Weng in February 2025 — within the open-source AI race, where it claims Inkling performs at a level comparable to leading open-weight models from China. WIRED frames the open-weight approac