Inkling

Thinking Machines Lab, the startup Mira Murati founded after leaving OpenAI, shipped its first from-scratch model on July 15, 2026 and put the full weights on Hugging Face under Apache 2.0. Inkling is a 975B parameter mixture-of-experts network with 41B active per token that accepts text, images, and audio, and it debuted as the leading US open-weights model at 41 on the Artificial Analysis Intelligence Index. Context reaches 1M tokens when self-hosting the raw weights, while the company's Tinker fine-tuning platform serves it at 64K and 256K windows with day-one support for post-training, the release being explicitly framed as a base for customization rather than a finished chat product. A smaller sibling, Inkling-Small, followed at the end of July as a 276B MoE with 12B active that matches the flagship on several reasoning and agentic benchmarks at a quarter of the size. Running the full model locally requires a multi-GPU cluster; the permissive license allows unrestricted commercial use, modification, and redistribution.