Liquid’s LFM3 puts multimodal AI in a smaller package—but deployment costs extend beyond the weights

Liquid AI has introduced LFM3, a family of small models intended to move more AI work onto the hardware running the application. Announced on 27 October, the release spans text, vision, audio, combined modalities and structured decision tasks. The commercial proposition is not simply a smaller chatbot: it is an application that can perform useful inference without sending every request to a remote model service.[1]

The family contains six models around the two-billion-parameter scale. Liquid says its new generation used 22 trillion pretraining tokens and supports adaptation on a single GPU. Those are the developer’s specifications and claims, not an independent finding that every model or fine-tuning workload fits every consumer machine.[1]

A long context is a capacity, not a speed promise

The text model’s card lists 1.73 billion parameters and a 256,000-token combined input-and-output context. Its 24 layers mix short convolutions, sliding-window attention and global attention. This is not a uniformly smaller copy of a conventional all-attention model; the architecture assigns different operations to different parts of the computation.[2]

The same card explicitly discourages using this model for agentic coding and complex multistep reasoning. That makes task selection essential: a compact model can be attractive for an embedded workflow without being an appropriate replacement for every larger reasoning system.[2]

Open weights, with a commercial threshold

Liquid’s licence allows substantial freedom to use and adapt the models, but commercial rights under the open licence end when a company’s annual revenue exceeds $10 million. A separate commercial agreement is then required. The terms contain an exception for qualifying research and nonprofit use. This is not the same commercial arrangement as an unrestricted permissive software licence.[3]

The distinction matters for procurement. A small project inside a large business is not necessarily a small licensee. Teams should evaluate the organisation to which the terms apply rather than assuming a low-volume application falls below a model-use threshold. The published condition is linked to company revenue, not a quoted allowance of tokens.[3]

The economic comparison should consequently be made at application level. Local deployment replaces a cloud inference bill with a combination of hardware capacity, maintenance, energy, integration and potentially a licence agreement. Whether that is preferable depends on how often the model runs, the acceptable delay and the value of keeping inputs within the device or organisation.

A sensible pilot would hold the task constant: the same documents, the same acceptable output and a defined latency limit. Measure failures as well as throughput. A cheap response that triggers manual repair can cost more than a dearer response that completes the job.

LFM3 widens the set of tasks worth testing locally. Its most interesting feature is the opportunity to choose a smaller, purpose-fit system—not a claim that all intelligence now fits into a one-gigabyte application.

Report an error or suggest a correction