Liquid AI has released LFM3, the third generation of its Liquid Foundation Models, and the pitch is not “a bigger model” but “six small models that share one brain.” A single 1.7B backbone was pre-trained once and then branched into a text model, a vision-language model, an audio model, an omni model and two decision models — all in the 2B parameter class, all post-trained for a different job.
The design premise is specialization on a single GPU. Every variant fine-tunes on one card, so a team can adapt a model to its own terminology and tooling and then run it on the device where the application lives. The text model fits in roughly 1 GB at 4-bit precision and carries a 256K context window, which is the number that matters for agents running locally.
Architecture by Arithmetic
LFM3-1.7B-Base has 24 layers in a tri-hybrid mix: 13 gated short convolutions (kernel size 7, up from 3 in LFM2), 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. On top sit head-wise gated attention, where a sigmoid gate can silence a head that has nothing to add — reportedly helping suppress activation outliers that hurt quantization — plus partial RoPE on global layers and sandwich normalization on both ends of every block.
The architecture was not chosen in the abstract. Liquid AI ran a hardware-in-the-loop search across 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. Base pretraining used 22T tokens; the final 3.5T-token mid-training stage leaned on curated data with a heavier share of agentic traces and tool calls, which is also what stretched context to 256K.
The Benchmarks
On agentic work the gains are lopsided: LFM3-1.7B scores 38.42 on WebArena-Verified against LFM2.5-2.6B’s 3.02, while using 35% fewer parameters, and it beats Qwen3.5-9B on long browser and business-app tasks. The vision branch, LFM3-VL-2B (1.9B params vs 3.1B for LFM2.5-VL-3B), lifts the average across 62 vision benchmarks from 57.6 to 71.4 — on par with Gemma 4 12B. Its new 164M-parameter vision encoder matches Qwen3.5-2B’s encoder with 2.2x fewer FLOPs.
On the audio side, the fully causal encoder can prefill while the user speaks, with time to first token as low as 50 ms, and LiquidCodec synthesizes speech at about 87x real time on a single CPU thread of an Apple M4 Max. LFM3-Audio-2B reaches UTMOSv2 3.83 versus 3.22 for its predecessor, across 18 languages. LFM3-Omni-2B, Liquid’s first omni model, beats models up to 10.7B parameters on Daily-Omni, WorldSense and AV-SpeakerBench. The decision models answer many structured questions in one forward pass with calibrated probabilities — d1-2B scores 75.5 against 64.0 for d1-3B on ImajevBench.
Everything ships through transformers, with a Colab notebook per fine-tuning recipe (SFT, DPO, GRPO, continued pretraining).
Related: LFM3 on Hugging Face, the LFM3-1.7B model card and Liquid AI’s agent harness guide