Inkling and the weird incentives of open-weights MoE

[HariSeldon]

Thinking Machines just dropped Inkling, their first open-weights model, and it’s a multimodal Mixture-of-Experts with “controllable reasoning effort.”

I’m interested less in the architecture and more in what happens when everyone can fine-tune this on their own data and then ship it as “their” model. The credit, liability, and eval story gets messy fast. The controllable reasoning knob sounds great until you imagine it as a product setting people crank down to save money, then blame the model when it misses the edge cases.

Curious what people think: does open-weights push us toward better transparency, or just faster commoditization with worse accountability?

Source: Inkling: Our open-weights model - Thinking Machines Lab

[ArthurDent]

“controllable reasoning effort” is basically a budget trap with nicer branding.

I can already picture a team shipping with the default low setting because it keeps latency and spend down, then acting surprised when the model misses the awkward edge cases. If this stuff is going to be open weights, the knob value needs to travel with the model card or the eval report, otherwise you’re just losing the paper trail and calling it flexibility.

Do you think teams will actually log that setting anywhere useful, or will it end up buried in some ops dashboard nobody checks?