Liquid AI has released LFM3, the new generation of its Liquid Foundation Models. This is a family of six open weight, class 2B models, around two billion parameters (the internal settings that determine a model’s behavior), built on a common base. They cover text, vision, audio, “omni” (which combines multiple modalities), and decision tasks.
A family of six models
All six models start from LFM3-1.7B-Base. The LFM3-1.7B model, with 1.73 billion parameters and a context window of 256K tokens (the amount of text a model can process at once, expressed in word pieces), is intended for local agents and has been post-trained for environments like Hermes Agent, Pi, and OpenClaw. LFM3-VL-2B is presented as the first vision model in the LFM family to accept video, sampled at 2 frames per second. LFM3-Audio-2B processes speech input and output in 18 languages, with tool calls. LFM3-Omni-2B combines vision, audio, and language, and responds in text or speech.
Finally, d1-2B and d1-Omni-2B are decision models that answer many questions in a single forward pass, the computation step that produces a response.
A hybrid architecture
The foundation combines three types of operators over 24 layers: 13 short gated convolutions, 5 sliding window attention layers, and 6 global attention layers. According to Liquid AI, this combination was found through a “hardware-in-the-loop” architecture search, meaning it takes real hardware into account. LFM2 was tuned on a phone and a laptop, while LFM3 was tuned on 14 device and backend combinations, including a Raspberry Pi 5 and the Qualcomm Hexagon NPU.
The base model was trained on 22T tokens, including the last 3.5T tokens oriented towards agents, which also increased the context to 256K. The post-training took place in four steps: supervised fine-tuning (SFT) on 1.75T tokens, mixed reinforcement learning, distillation of six “specialist” models into one, then agentic reinforcement learning.
Results announced by the publisher
According to Liquid AI’s evaluations, LFM3-1.7B achieves an average of 49.37 on 29 benchmarks, compared to 47.24 for Qwen3.5-9B. On WebArena-Verified, it scores 38.42 versus 28.94 for Qwen3.5-9B, and on BFCL V4, 71.39 versus 62.83. However, the model lags behind on agentic knowledge and code: GPQA Diamond shows 63.37 versus 80.40 for Qwen3.5-9B, and SWE-bench Verified 43.40 versus 51.20. On Humanity’s Last Exam, the score rises from 7.17 to 13.90 with tools.
For vision, LFM3-VL-2B displays an average of 71.4 on 62 benchmarks, compared to 71.7 for Gemma 4 12B on the same suite. LFM3-Audio-2B achieves a word error rate (WER) of 6.2 in English and a score of 3.8 on UTMOSv2 for speech synthesis. LFM3-Omni-2B scores 70.1 on Daily-Omni, versus 61.3 for Qwen2.5-Omni-7B.
The LFM Open License v1.0 allows free commercial use up to 10 million dollars in annual revenue.
Deployment
The weights are available on Hugging Face, on the Liquid AI Playground, and via the Liquid API, with GGUF, MLX, and ONNX formats for text and vision models. The cited runtime environments include llama.cpp, MLX, vLLM, SGLang, Ollama, and Transformers. Liquid AI indicates that in 4-bit, LFM3-1.7B requires about 1 GB of memory, making it suitable for phones and laptops.
On an NVIDIA B300 with SGLang, the model produces 3.6B tokens per day for 1,024 concurrent requests. Liquid AI also shows that specialization is possible: a few hours of supervised fine-tuning and GRPO learning on a single A100 raised wind turbine defect classification to 67%.
