I've got an Inkling

Thinking Machines put out its first open-weights model this week, and it's called Inkling. The headline spec is a 1 million token context window, and the model handles text, images and audio rather than just text alone. For a lab that's been mostly known for its research work, shipping something you can actually download and run is a real shift.

Here's the catch, though: Inkling isn't close to the best open models available right now. Most of those top performers come out of Chinese labs these days, and GLM-5.2 in particular has become the model a lot of startups are quietly switching to, running it self-hosted or fine-tuned instead of paying for frontier API access. GLM-5.2 does have a gap of its own — no vision support, so you can't drop an image into a prompt — which is exactly the kind of hole third-party tooling has started filling.

So why release a model that isn't chasing the top of the leaderboard? Because Inkling isn't really meant to win on raw capability out of the box. It's built to live on Tinker, Thinking Machines' own fine-tuning platform, where the point is turning a decent base model into something sharper for a narrow job. That's a different bet than trying to out-benchmark everyone else, and it's a bet on a market that's growing fast: companies increasingly want a model they can shape around their own data and workflows rather than a generic one they rent by the token.

And that framing matters more than the raw scores might suggest. If the real competition is shifting toward who has the best fine-tuning pipeline rather than who has the single smartest base model, then a merely decent multimodal model with a huge context window and a home platform built around customization could carve out a lane of its own, even while sitting behind the leaders on paper.