Liquid AI ships LFM3, a 1.7 billion parameter model that beats its larger predecessor

What happened

On October 27, 2026, Liquid AI published LFM3, a family of open-weight models built from LFM3-1.7B-Base. The base has 24 layers and was pretrained on 22 trillion tokens, with the last 3.5 trillion used to extend the context window to 256,000 tokens. Six models share that backbone: text, vision with video, audio across 18 languages, an omni model, and two decision models that return a probability instead of a paragraph. Liquid says LFM3-1.7B outperforms LFM2.5-2.6B with 35 percent fewer parameters. On WebArena-Verified the score is 38.42, against 3.02 for the predecessor. Liquid also says it beats Qwen3.5-9B on long browser and business-app tasks, and does not print that model's score. The vision model, 1.9 billion parameters against 3.1 billion for LFM2.5-VL-3B, raises Liquid's average across 62 vision benchmarks from 57.6 to 71.4, which Liquid says is on par with Gemma 4 12B. The audio speech-quality score UTMOSv2 moves from 3.22 to 3.83. A decision model, d1-2B, scores 75.5 against 64.0 on ImajevBench.

Why it matters

The useful change is that the small model is the one with the higher score, and the weights are downloadable today. A 1.7 billion parameter file that Liquid can fine-tune on one GPU is a different object from a hosted frontier model. The scores that justify the swap are still Liquid's, including a WebArena jump from 3.02 to 38.42 that no one else has rerun.

Liquid says a fine-tune of the vision model, trained across 15 robot embodiments, reaches 96.2 percent average success on the four LIBERO suites and 45.7 percent on RoboCasa-GR1, and generates action chunks in under 50 milliseconds on a Jetson AGX Orin. Those figures are for the fine-tune, not the base checkpoint.

The new vision encoder is 164 million parameters. Liquid says it uses 3 times less compute than the SigLIP2 encoder in earlier LFM vision models, and 2.2 times fewer operations than Qwen3.5-2B's vision encoder when both are measured with the same small language backbone.

LiquidCodec turns generated tokens into speech at about 87 times real time on one CPU thread of an Apple M4 Max. Time to first audio token is as low as 50 milliseconds, on Liquid's account.