Inkling, an open-weight model designed to be adapted, not to top rankings

Inkling is not meant to be the most powerful AI model on the market, but one that is easy to take and reshape: Thinking Machines Lab presents it as a base other companies can download and adapt to their own uses. The company releases the full weights of the already-trained model -what the industry calls "open weights"-, so anyone can download it, run it on their own hardware and modify it with their own data, without depending on a closed service. It is not the same as traditional open source: the company did not release the training data or process, but it did release the final model.

The model natively processes text, images and audio. It is large, with 975 billion total parameters, though only 41 billion are active for each response thanks to an architecture that splits the work across several specialised subsystems, which reduces the cost of each query. It supports context windows of up to one million tokens and was trained on 45 trillion tokens of text, images, audio and video.

It lets users adjust how much time it spends reasoning before answering: more effort brings better results, but also more cost and waiting time. According to the company's own tests, it needs fewer resources than other open models to reach similar performance; in one coding test, it matches Nemotron 3 Ultra using roughly a third of the tokens. In a comparison of apps generated by several models and rated by people, it ranks among the best-regarded open-weight models, though behind the most advanced closed systems.

Alongside Inkling, the company previewed Inkling-Small, a lighter version with similar results on several tests, built for tasks where speed matters more than raw power. The model can already be fine-tuned on Tinker, with a limited-time 50% discount, and downloaded from Hugging Face. The company also commissioned external safety testing, in which Inkling refused harmful requests better than comparable models, without blocking legitimate queries.

Key points

• Thinking Machines Lab's first open-weight model, built to be customized rather than to top rankings
• Full weights released, but not the training data or process: not traditional open source
• Natively processes text, images and audio
• 975 billion total parameters, only 41 billion active per response
• Context window up to one million tokens; trained on 45 trillion tokens
• Adjustable reasoning effort to balance cost and performance
• Matches Nemotron 3 Ultra on a coding test using roughly a third of the tokens
• Ranks among the best-regarded open-weight models in human-rated comparisons, behind top closed systems
• Released alongside Inkling-Small, a lighter, faster preview version
• Passed external safety testing, refusing harmful requests without blocking legitimate ones