Liquid AI did not ship a single checkpoint this week. It shipped a family. The LFM3 release pairs six open-weight systems in the 2B range, each tunable on one graphics card and small enough to run on a handset, a laptop, a private server or in the cloud, with no custom silicon required at either end of that range.
All six descend from a shared trunk. LFM3-1.7B packs 1.73B parameters behind a 256K window and was post-trained for agent harnesses. LFM3-VL-2B is the lab’s first vision system, sampling video at 2 frames per second. LFM3-Audio-2B takes speech in and returns it across 18 languages with tool calls, LFM3-Omni-2B merges all three modes, and d1-2B plus d1-Omni-2B are decision models that settle many queries during one forward pass. Weights arrive in GGUF, MLX, ONNX, FP8 and NVFP4, with 4-bit builds near 1 GB of memory.
The trunk interleaves three operator types over 24 layers: 13 gated short convolutions, five sliding-window attention blocks and six global attention blocks. Liquid arrived at that recipe through hardware-in-the-loop architecture search spanning 14 device and backend pairings, among them a Raspberry Pi 5 and a Qualcomm Hexagon NPU.
Agent benchmarks favor the little model. WebArena-Verified gives the 1.7B a 38.42 against 28.94 for Qwen3.5-9B, and BFCL V4 yields 71.39 versus 62.83. Knowledge still rewards size: GPQA Diamond lands at 63.37 against 75.76 for Qwen3.5-4B, and SWE-bench Verified at 43.40 versus 51.20.
The license is open-ish rather than open, and the terms shift once a company gets big. LFM Open License v1.0 sets a $10M revenue threshold, so commercial users beyond that scale should expect a different conversation with the lab before deployment.
