A small model can make local deployment practical, but the download size is only the first part of the hardware calculation.
Liquid AI released LFM3 on October 27, introducing a new family of small models for specialized tasks. The company says the first release includes six models built from LFM3-1.7B-Base, covering text, vision, audio, omni, and decision workloads.
A shared backbone with different jobs
The release announcement places the family in the roughly 2-billion-parameter class. Liquid says base and post-trained weights are available on Hugging Face, along with access through its Playground and API.
Liquid reports a 256K context window for the text model and says its 4-bit weights fit in about 1 GB. These are the company’s specifications and measurements. We have not reproduced its benchmarks or run the models locally.
Calculate weights separately from working memory
A simple calculation helps interpret the text-model claim: 1.7 billion parameters multiplied by 4 bits, divided by 8, equals 850 million bytes. That is about 0.85 GB in decimal units before quantization metadata and other overhead.
Serving the model also needs runtime memory, temporary buffers, and context state. A supported long context can therefore demand more memory than the weight file alone. Input modality, runtime, precision, and the number of simultaneous requests also affect the deployment.
Before choosing a device, measure a representative request at the context length you intend to use. Record peak memory and response time, and assess whether the answer meets the task. A short demonstration cannot establish a long-context hardware budget.
Open weights come with a commercial boundary
The pinned LFM3-1.7B license uses LFM1.0 terms. It sets an annual-revenue threshold of $10 million and conditions commercial-use rights on that limitation. It also contains a qualified nonprofit provision for noncommercial or research uses.
This is not an Apache 2.0 release. A company evaluating deployment should check the actual terms against its organization and use case before treating downloadable weights as unrestricted commercial permission.
Choose the model for a narrow task
Start with one job with a clear expected result, such as classifying a small set of local documents. Choose the required modality first, then compare the result with the memory and license constraints.
Our earlier LFM2.5 deployment guide covers a different model generation. Read Liquid’s model-family walkthrough for the new release details.
