This week, Liquid AI open-sourced d1, a multimodal decision model on Hugging Face designed for edge (on-device) inference. We see this as the open-source camp formally adopting "runnable on-device" as a product position, rather than continuing to scale parameters. The company is an MIT spinout specializing in what it calls Liquid Neural Networks — a sparser, lower-power architecture.
What this is
d1 is a multimodal model supporting text and image input (multimodal = processing multiple input types simultaneously, like text, images, audio), built for local decision tasks — robotic control, in-vehicle interaction, industrial visual inspection. A "decision model" means it doesn't just answer questions — it outputs the next action or control command directly. Compared with traditional large models, its engineering significance lies in fusing two previously contradictory things: "understanding the world" and "not needing the internet."
Industry view
Hugging Face co-founder Clem Delangue called it "an important step for the open-source ecosystem" on X. We believe that for industries with strict data compliance requirements — finance, healthcare, manufacturing — local inference means lower compliance costs and a further reduction in the bar to AI pilots.
But we note two counterpoints. First, a SambaNova engineer pointed out that decision-class tasks demand extremely high real-time performance — open-sourcing alone isn't enough; companion hardware accelerators are required to be useful. Second, "open source" doesn't mean "commercially usable" — d1 ships under a research license that explicitly restricts commercial use; enterprises looking to deploy will have to wait for a commercial release.
Impact on regular people
For enterprise IT: AI pilot barriers in finance, healthcare, and manufacturing will drop further; projects previously stalled by data compliance may restart.
For individual careers: Within the next 1-2 years, on-phone AI assistants may handle 80% of tasks locally — no need to upload every sentence to the cloud.
For the consumer market: Smart home, automotive, and wearable devices will get smarter and more power-efficient — but in the short term, there's still no "killer app" compelling users to upgrade their hardware.