An October 27 launch, covered as catch-up
Liquid AI announced LFM3 on October 27. Our October 28 briefing catches up on that release. Liquid describes a family spanning several modalities, but LFM3-1.7B itself is a text-only checkpoint. [1][2]
Its model card lists 1.73 billion parameters and a 256K-token context window covering input and output together. The intended work includes local agents, tool calling, structured output, and specialization. Liquid explicitly discourages agentic coding and complex multi-step reasoning for this small model. [2]
A small estimate is not a complete memory budget
The BF16 checkpoint's repository lists model.safetensors at 3.46 GB. Liquid estimates about 1 GB for the model at four-bit precision; that is a different representation from the listed checkpoint and does not establish total application memory. No weights were downloaded during our review. [2][3]
Context storage, the inference engine, tools, and other running applications also need resources. Treat the advertised context capacity as a limit to assess, rather than proof that every device can use the full window reliably.
Read the custom license
The checkpoint uses LFM Open License v1.0. It has conditions on commercial use, including a revenue-related limitation, and redistribution requirements. It is not unrestricted Apache 2.0. Read the actual license for the intended entity and use; downloadable weights do not settle commercial eligibility. [2][4]
Our take
For technicians, start with a narrow task and representative examples. Measure complete memory use and check generated tool arguments before granting access. Local inference can be useful without replacing a larger reasoning model, and connected tools can still send data outside the device.
We reviewed public sources; we have not run the model or verified its performance.