Liquid AI has released LFM3, its new generation of Liquid Foundation Models. Liquid AI LFM3 is a family of 6 open-weight, 2B-class models built on one shared backbone. The family covers text, vision, audio, omni and decision tasks. Every model can be fine-tuned on a single GPU. Developers can then deploy on phones, laptops, private servers or the cloud.
TL;DR
- Size: LFM3-1.7B has 1.73B parameters and a 256K context window. LFM3-VL-2B has 1.92B and 128K. LFM3-Omni-2B has 2.2B.
- Runs on: about 1 GB of memory at 4-bit, so phones and laptops. GGUF, MLX, ONNX, FP8 and NVFP4 checkpoints ship.
- Performance: LFM3-1.7B averages 49.37 across 29 benchmarks, above Qwen3.5-9B at 47.24 (Liquid’s own eval).
- Best: IFEval 96.66, and AutomationBench 41.29 vs 15.10 for Qwen3.5-9B.
- Worst: Humanity’s Last Exam (no tools) 7.17, and RULER 128K 65.56 vs 87.11 for Qwen3.5-9B.
What is Liquid AI LFM3?
LFM3 is a multimodal model family from Liquid AI that branches 6 models from 1 base. All of them start from LFM3-1.7B-Base.
The release includes:
- LFM3-1.7B: a text model for local agents, post-trained for harnesses like Hermes Agent, Pi and OpenClaw.
- LFM3-VL-2B: the first LFM vision model with video input, sampled at 2 fps.
- LFM3-Audio-2B: speech in and out across 18 languages, with tool calling.
- LFM3-Omni-2B: vision, audio and language in one model, answering in text or speech.
- d1-2B and d1-Omni-2B: decision models that answer many questions in 1 forward pass.
How does the tri-hybrid architecture work?
The backbone mixes 3 operators across 24 layers. There are 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. The SWA window is 256 tokens. Convolution kernels grew from 3 to 7.
Liquid AI team found this mix with its hardware-in-the-loop architecture search. LFM2 was tuned on 1 phone and 1 laptop. LFM3 was tuned across 14 device and backend combinations. These include a Raspberry Pi 5 and the Qualcomm Hexagon NPU.
Other changes include head-wise sigmoid gates on attention, partial RoPE on global layers and sandwich norm. The base model saw 22T tokens. The last 3.5T tokens were agent-heavy mid-training, which also extended context to 256K.
Post-training ran in 4 stages. SFT on 1.75T tokens came first, then mixed RL. Next, 6 specialist teachers were distilled into 1 model, followed by agentic RL.
How does LFM3-1.7B perform on benchmarks?
Liquid AI reports that LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters. The bigger story is the agent results. On WebArena-Verified it scores 38.42, against 28.94 for Qwen3.5-9B. On BFCL V4 it scores 71.39, against 62.83.
Size still wins on knowledge and agentic coding. GPQA Diamond is 63.37, versus 80.40 for Qwen3.5-9B. SWE-bench Verified is 43.40, versus 51.20. Search narrows the gap: Humanity’s Last Exam rises from 7.17 to 13.90 with tools.
