Thinking Machines Reveals Inkling: What a 975B Open-Weights Model Means

What Was Released

Thinking Machines Lab revealed Inkling, its first in-house open-weights model, in an official announcement on July 15. According to the company, Inkling is a Mixture-of-Experts model with 975B total parameters, while using about 41B active parameters during actual inference.

Put simply, the overall warehouse is very large, but the model pulls out only the necessary workbenches for each question. That makes it an approach aimed at balancing the broad capabilities of a large model with inference cost.

What Open Weights Means

The most important term in this announcement is open-weights. Unlike closed models that are used only through an API, an open-weights model lets outside developers and organizations obtain the model files, test them in their own environments, or modify them. Inkling also has a page on Hugging Face, confirming that the release was not limited to a simple announcement post.

This approach gives developers more freedom. For example, it leaves more room to tune a model around the way each organization works, such as internal document formats, specific coding styles, or product-specific response rules.

Why Tinker and a 1 Million-Token Context Matter

Thinking Machines did not introduce Inkling as a standalone model alone. It connected the release with the company’s customization platform, Tinker, along with a model card, documentation, and example workflows. The announcement says Inkling supports a context of up to 1 million tokens.

A 1 million-token context points toward handling long codebases, large research materials, and lengthy planning documents more broadly at once. In Game Sunakku terms, it is closer to an AI answering with an entire drawer of thick work notes spread out in front of it, rather than looking only at a short memo.

What to Watch Carefully

That said, it would be an exaggeration to read this announcement only as “the strongest model has arrived.” Thinking Machines does not say Inkling is the strongest model in every area. Instead, it emphasizes Inkling as a customizable starting point.

Open-weights models bring responsibility along with their freedom. Organizations that adopt them need to check fine-tuning, deployment cost, data security, safety evaluation, and license terms for themselves. In particular, when a company adapts a model to its own work, it also needs to verify directly whether the resulting system is safe and accurate.

Game Sunakku Takeaway

The release of Inkling points in a somewhat different direction from the API-centered competition led by companies such as OpenAI, Anthropic, and Google. The key point is not “another famous chatbot has appeared,” but that competition around organizations directly refining and using large models is expanding again.

For developer readers, three points are worth checking before focusing on a single line in a benchmark table. First, the actual download and runtime conditions. Second, how easy customization tools such as Tinker are to use. Third, whether the long context and multimodal training base are genuinely useful for your own work. Only when those three points line up does Inkling become more than news and turn into a real tool candidate.