For most of the AI boom, the most capable models have lived behind APIs — you could talk to them, but you could not see, modify or privately host their inner weights. That wall moved this month when Thinking Machines Lab released Inkling, its first model trained from scratch and published with fully open weights.
The headline numbers are large but worth understanding. Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters, yet only about 41 billion are "active" for any single request. MoE is the trick behind many modern giants: instead of running the whole network every time, the model routes each piece of input to a small subset of specialized sub-networks, or "experts." The result is a model with a huge capacity for knowledge that still runs at a fraction of the cost of a same-sized dense model.
Inkling was pretrained on roughly 45 trillion tokens spanning text, images, audio and video, which is why it is described as natively multimodal — it does not bolt on separate vision or speech modules after the fact. It also supports a context window of up to one million tokens, enough to hold an entire book or a long technical archive in a single conversation.
The part that matters most for builders is control. Because the weights are open, anyone can download Inkling, run it on their own hardware, and fine-tune it for a specific task using the lab's "Tinker" tooling. The model also exposes adjustable "thinking effort," letting developers trade reasoning depth against latency and token cost depending on the job. That combination — open, multimodal, long-context and tunable — is what sets it apart from closed frontier systems and from smaller open models that lack its scale.
There is a wider story here. Open-weights releases shift power from a handful of labs toward the broader research and startup community. They let hospitals, universities and companies build private, specialized assistants without sending sensitive data to a third party. The trade-off is responsibility: an open model can be used by anyone, for any purpose, which is precisely why the labs publishing them emphasize safe fine-tuning and usage guidance.
Knowledge takeaway: Inkling is a 975B-parameter open-weights Mixture-of-Experts model from Thinking Machines Lab that activates ~41B parameters per query, handles text, image and audio natively, and supports a 1M-token context; by publishing its weights it lets anyone privately host and fine-tune a frontier-scale model, moving capability out of closed APIs and into the hands of the wider community.