Intelligence where
your data already is.
Five working systems. Each runs a fine-tuned Liquid model on a phone with the network off, an edge GPU in front of your traffic, or a VPC you control.
The five systems
Each one is a specialist that does a single job inside a pipeline, next to whatever frontier model you already use. The full model roster is at liquid.ai/models.
Search Systems
Nobody types the words that are in your catalog. The pipeline rewrites the query before it reaches the index, and rescues the searches that would have come back empty.
Pipeline flagshipOne 350M specialist and a token-level retriever, on one edge GPU
Orchestration & workflow
Agents you can actually let run. Every tool call gets checked before it executes.
Pipeline flagshipLFM2.5-350M pre-flight check, on one edge GPU
Vertical foundation models
One frozen backbone per industry. Each new business question becomes an adapter on top of it rather than a model of its own.
Architecture flagshipA frozen 350M backbone with a small encoder and task heads
Recommendation systems
The model writes recommendations straight from your catalog. Everything it names is something you actually stock.
Split-screen flagshipSame host, same prompt, same catalog
When a specialist model or agent is the right answer
- The inference bill is growing faster than the AI budget.
- Product wants this on the device, and the ML team says not with current models.
- A feature shipped, and latency is dragging the experience down.
- Compliance moved the line to on-prem, in-region, or fully disconnected.
- A working prototype now has to handle a hundred times the volume.
- A hardware refresh locked the RAM budget before the model was chosen.
How one of these gets made
- 01The examples
A dozen real cases, plus the rules a team already applies by hand.
- 02The data
Growing those into a representative set that covers the edge cases. Most of the time goes here. The training run takes hours on one GPU, scored against a held-out set.
- 03The placement
Model, quantization, and serving engine chosen together for a phone, an edge GPU, or your VPC. We profile on your hardware, not ours — the same efficiency, measured on the silicon you are shipping.