Murati's lab releases its first open model

Mira Murati’s Thinking Machines Lab put out its first proprietary model on July 15, called Inkling, and broke from the pattern OpenAI, Anthropic, and Google have set with their flagships by shipping it with open weights. Anyone can pull it from Hugging Face and modify it directly.

Inkling is a mixture-of-experts model with 975 billion total parameters, activating roughly 41 billion of them per task, the usual way labs keep enormous models affordable to serve. Training ran across 45 trillion tokens spanning text, images, audio, and video; the model reasons natively across all of them and handles context windows up to 1 million tokens. Reported benchmark scores: 77.6% on SWE-bench Verified, 97.1% on AIME 2026, and 87.9% on GPQA Diamond. A smaller preview variant, Inkling-Small, runs at 12 billion active parameters.

The model is live now for fine-tuning on Tinker, Thinking Machines’ customization platform, at a 50% introductory discount, and is also available through Together, Fireworks, Modal, Databricks, and Baseten.

The strategic bet is explicit: enterprises may care less about renting the single smartest general-purpose model than about owning one they can shape to their own data and workflows. That is a different wager than the one OpenAI, Anthropic, and Google are making with closed, subscription-priced flagships, and it puts Thinking Machines in more direct competition with open labs like Meta, Mistral, and China’s DeepSeek and Zhipu.

For operators evaluating whether to rent or build, Inkling is now a concrete data point: a frontier-scale, multimodal, open-weight model with a fine-tuning platform attached, from a lab with the funding and pedigree to keep iterating on it. Related reading: see what open models good enough for operators means for your stack, and how to think about routing tasks between AI models as more open options like this one arrive.