I think "embodied" is a word with decades of past associations in cognitive science and AI and if you're claiming that your model is embodied you should try to back that up with specific evidence.
This seems cool but I don't think they live up to their self-description
It wasn’t that long ago that we were experimenting on with llama derived models. There is still a lot happening with open weight models, but you may have to look at those models meant for targeting edge devices. In my own opinion, small models that can do useful work are going to become a much bigger deal in domains where data privacy is a requirement. Most businesses were technology is not the core mission do not want to host their own frontier models.
Whether one needs a custom frontier model or tuned model or an existing one depends on the task at hand. Custom tasks that are not covered by existing models will inevitably require a new frontier model independent of whether the business is a technology firm or not. An LLM or a basic vision model isn't suitable for all tasks. For example, imagine a model that outputs coherent architecturally sound designs for bridges or electrical systems that a mere rule-based system cannot output.
I work in this space and do a lot of experimentation with AI. We are still a very long way from AI being able to go from owner requirements to finished design or even an acceptable solution. In part that’s because the work products are locked up in proprietary formats and are not easily understood by these models. The industry conventions are based on extremely slow, incremental improvements on Penn and paper based processes that are about a century old. What may surprise you is that computational design is a highly under explored space for market-based reasons that are so complex I would need to write a lengthy blog post just to explain it. If large language models disappeared, you could still have almost entirely automated structural design for Bridges. A large amount of design is just the application of well defined rules to a site specific problem. There are characteristics in shortcuts people take, but you should always be able to get to the same result from first principles.