Thinking Machines Lab releases Inkling, an open-weight 975B model

Background

The company published its reasoning for releasing the model with open weights in a separate essay ^[2].

What happened

On 15 July 2026 Thinking Machines Lab released Inkling, its first model, with open weights ^[1]. It is a mixture-of-experts model with 975 billion total parameters and 41 billion active parameters ^[1]. In a mixture-of-experts model, only part of the network runs for each piece of input, which keeps the cost of each step lower than the total size would suggest.

Inkling handles text, images and audio, and supports up to 1 million tokens of context ^[1]. The context length is the amount of text the model can consider at one time.

Key details

The company reported 77.6% on SWE-Bench Verified, a test of software engineering tasks, and 87.2% on GPQA Diamond, a set of hard science questions ^[1]. Benchmark results show how a model performed on fixed tests, and they do not settle how useful it is in other settings.

A smaller model, Inkling-Small, with 276 billion total parameters and 12 billion active, followed two weeks later ^[1]. Offering a range of sizes lets users choose a model that fits their hardware and budget.

The company published its reasoning for open release in an essay called "A Safe Path to Open Weights" on 31 July 2026 ^[2]. The same day, the smaller model was released. Publishing the reasoning alongside the models allowed readers to see the company's stated case for open weights.

What open weights allow

Open weights let researchers and companies study a model, test it on their own data, and fine-tune it for their own needs. They also mean that the release is not controlled by a single hosted service. The trade-off is that the developer no longer controls every way the model is used once the weights are public, which is why the company published its reasoning separately ^[2].

Why it matters

The release is notable for its size and for the written explanation published alongside the weights. It shows a model-building lab choosing to publish both. Readers who want to follow open models can compare the reported results with later independent tests.