Liquid AI releases its third-generation Liquid Foundation Model, LFM3: six 2B-class models share the LFM3-1.7B-Base backbone, covering text, vision (with first-time video input support), speech, omni-modal, and "decision" outputs. The weights are available on Hugging Face, targeting local execution on devices such as smartphones and laptops, with single-GPU fine-tuning.
Confidence: High (80) · Event time 2026-10-27
Reasoning
- The company states that the new tri-hybrid architecture (gated short convolutions + sliding window attention + global attention) allows LFM3-1.7B to outperform LFM2.5-2.6B with 35% fewer parameters. The largest improvements are seen in agentic tasks: WebArena-Verified scores reached 38.42 (previous generation: 3.02). The text model, after 4-bit quantization, is approximately 1GB, with a 256K context window and 22T tokens of pretraining.
- LFM3-VL-2B improved its average score across 62 vision benchmarks from 57.6 (LFM2.5-VL-3B) to 71.4, with the company claiming it is comparable to Gemma 4 12B. The custom vision encoder has one-third the computational cost of SigLIP2.
- LFM3-Omni-2B is Liquid's first omni-modal model, capable of seeing, hearing, and responding via speech, potentially replacing the pipeline of speech recognition + vision-language model + LLM + text-to-speech. The audio encoder is fully causal, with the company claiming a first-token latency as low as 50 milliseconds.
- d1-2B and d1-omni-2B do not generate text; instead, they answer multiple questions about a single input in one forward pass, providing calibrated probabilities for each. These are intended for structured judgments such as ticket classification and invoice field extraction.
Usage
- The models are available under the LiquidAI organization on Hugging Face (LFM3 collection, including LFM3-1.7B-Base for custom post-training) and can be loaded using transformers (requires trust_remote_code); model cards also list inference engines such as llama.cpp.
- The company states that each model can be fine-tuned on a single GPU; LFM3-1.7B has been post-trained for agentic frameworks such as Hermes Agent, Pi, and OpenClaw.
Risks
- The license is the LFM Open License v1.0, not Apache or MIT: commercial use by companies with annual revenue of 1,000 million USD or more is not covered by the license.
- All benchmarks are self-reported by Liquid, with no third-party reproduction; actual model outputs have not been reviewed.