VarwoLiquid LFM3: Model Family, Downloads and Local Evaluation
Choose the LFM3 checkpoint for text, vision or audio, find official weights and distinguish provider claims from local test results.
LFM3 is Liquid AI's open-weight family for specialized applications, announced October 27, 2026. Start with the official announcement and the particular model card, rather than treating every LFM3 download as the same model. Sources checked October 31, 2026.
Choose a checkpoint
| Task | Model family to inspect |
|---|---|
| Text, tool calling, structured output | LFM3-1.7B |
| Images and video understanding | LFM3-VL-2B |
| Speech and audio tasks | LFM3-Audio-2B |
| Combined visual, audio and language input | LFM3-Omni-2B |
| Specialized decisions | d1 models in the release family |
The announcement describes six models around the 2B class. Its roughly 1 GB at 4-bit statement concerns text-model weights, not a guarantee that the entire application and long-context cache fit in 1 GB.
The LFM3-1.7B card lists 1.73B parameters, a 256K input-plus-output context and LFM Open License v1.0. Open weights do not by themselves mean an Apache or MIT licence. Read the checkpoint's licence before deployment. The card cautions against choosing this small text model for difficult agentic coding or complex multistep reasoning.
Plan a local check
Follow the linked card's current runtime instructions. Record the exact checkpoint, quantization, device, context length and runtime version. Try one representative task and inspect both correctness and memory use before increasing context or concurrency. Compare models under the same settings and keep failed tasks in the result.
Liquid's published benchmarks are provider evaluations. Varwo has not downloaded the weights, installed a runtime or reproduced those measurements. For another model-selection example, see the Kimi access and pricing guide.