Inkling: Our open-weights model

Here's the thing: 975 billion parameters sound impressive, but in practice, it raises immediate questions about operational burden. A Mixture-of-Experts model like Inkling can theoretically manage resource usage by activating only a fraction of its parameters, but that requires a robust infrastructure to handle the complexities of routing and managing which experts to activate. You could easily end up with a convoluted setup that requires more maintenance and oversight than you bargained for, especially when those performance benchmarks are still absent from the discussion.

What they're not saying: The claims of being multimodal and trained on 45 trillion tokens are enticing, yet without real-world application data, it's hard to gauge how Inkling performs under pressure. Training data volume doesn't guarantee effectiveness. If you're considering this for production, keep in mind the additional overhead associated with integrating a model of this size, especially in a pipeline that might already struggle with data quality.

To be clear: Inkling-Small, while still in testing, could present a more manageable alternative if it delivers on its promise. However, you should be cautious about adopting either variant until you have concrete performance metrics and understand how these models will interact with your existing tech stack.

In short, if you're eyeing this model, approach with caution. It may not be ready for prime time, and the complexities involved could lead to more headaches than solutions. If you’re not prepared to invest the time in understanding its operational quirks and potential pitfalls, you might want to sit this one out for now.