Thinking Machines ships Inkling, its 975B open-weight MoE

The interesting part of Mira Murati's first release from Thinking Machines is not that the model exists, which was inevitable after the $12 billion seed valuation, but the bet the release is placing. Inkling is a 975 billion parameter mixture-of-experts model, roughly 41 billion of which are active for any given token, and the company posted the full weights on Hugging Face under Apache 2.0. It accepts text, image and audio inputs and, according to launch coverage, was pretrained on 45 trillion tokens. That is a frontier-scale release from a lab that has never shipped a public model before.

What is unusual is what Thinking Machines is not claiming. They are not saying Inkling beats the closed frontier models. The company is upfront that Inkling is not the strongest model available, and is positioning it as a starting point meant to be customized. Their revenue engine is Tinker, the fine-tuning platform, which already serves customers like hedge fund Bridgewater Associates for financial tasks. Inkling gives Tinker something it did not have before, an in-house open base model designed to be reshaped rather than rented.

The reason that matters if you are a CTO or a research lead is that a serious new open-weight entrant, from an ex-OpenAI team that includes Murati, John Schulman and Lilian Weng, changes the shortlist for anyone who wants to own their base model rather than pay per-token indefinitely. The 41 billion active parameter design is small enough to be economically served, which is likely why the launch partners include Together AI, Fireworks, Modal, Databricks and Baseten for hosted inference.

The honest caveat is the one Thinking Machines volunteers itself. The benchmark numbers the model card publishes, including 97.1% on AIME 2026 and 77.6% on SWEBench Verified, are competitive but self-reported at what the card calls effort=0.99, and the company is clear this is not category-leading. What the coverage does not give you is a real cost-per-token comparison, or much detail on the training data mix beyond publicly available sources, third-party acquisitions and synthetic generation.

The forward-looking bet is that enterprises will end up caring less about which lab has the smartest model this quarter and more about which base model they can genuinely make their own. If that is right, this is the moment Thinking Machines actually enters the game.