Thinking Machines releases Inkling: a 975B-param open-weights MoE model with 41B active params

Inkling: Our Open-Weights Model

Thinking Machines open-sourced Inkling, a 975B-total-param, 41B-active MoE model with a 1M-token context window, pretrained on 45T multimodal tokens. It handles text, images, and audio natively, with controllable thinking effort, and is positioned as a customizable base rather than a benchmark leader. A 12B-active Inkling-Small preview was also announced. The team demoed self-fine-tuning on their Tinker platform: the model wrote a training job, ran it, and switched to the new weights in about 27 minutes. Weights are on Hugging Face; fine-tuning is available on Tinker.

Why it matters: Thinking Machines' first open-weights model: 975B MoE, native multimodal, 1M context, pitched as a hackable base rather than a benchmark leader—clear differentiation. Not p1 because it's a debut from a new team with no ecosystem or real-world validation yet; featured is the ri...