Liquid AI Releases LFM3, Six Open-Weight 2B-Class Multimodal Models on One Backbone

The release includes LFM3-1.7B, a text model for local agents post-trained for harnesses such as Hermes Agent, Pi and OpenClaw; LFM3-VL-2B, the first LFM vision model with video input sampled at 2 frames per second; LFM3-Audio-2B, which handles speech in and out across 18 languages with tool calling; LFM3-Omni-2B, which combines vision, audio and language and answers in text or speech; and d1-2B and d1-Omni-2B, decision models that answer many questions in one forward pass. All six start from LFM3-1.7B-Base.

LFM3-1.7B has 1.73 billion parameters and a 256K context window. LFM3-VL-2B has 1.92 billion parameters and 128K context, and LFM3-Omni-2B has 2.2 billion. Liquid AI says LFM3-1.7B runs in about 1 GB of memory at 4-bit, and it ships GGUF, MLX, ONNX, FP8 and NVFP4 checkpoints.

Liquid AI reports that LFM3-1.7B averages 49.37 across 29 benchmarks, above Qwen3.5-9B at 47.24 on the company's own evaluation. It cites IFEval at 96.66 and AutomationBench at 41.29 versus 15.10 for Qwen3.5-9B. The model is weaker on Humanity's Last Exam without tools at 7.17 and on RULER 128K at 65.56, against 87.11 for Qwen3.5-9B.

On agent benchmarks, Liquid AI reports WebArena-Verified at 38.42 for LFM3-1.7B versus 28.94 for Qwen3.5-9B, and BFCL V4 at 71.39 versus 62.83. Qwen3.5-9B remains ahead on knowledge and coding, with GPQA Diamond at 80.40 versus 63.37 and SWE-bench Verified at 51.20 versus 43.40. With tools, Humanity's Last Exam rises from 7.17 to 13.90 for LFM3-1.7B. Liquid AI also says LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters.

The backbone mixes three operators across 24 layers: 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers, with a 256-token sliding-window attention window. Convolution kernels grew from 3 to 7. Liquid AI found the mix through hardware-in-the-loop architecture search. LFM2 was tuned on one phone and one laptop; LFM3 was tuned across 14 device and backend combinations, including 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 22 trillion tokens, including a final 3.5 trillion tokens of agent-heavy mid-training that extended context to 256K. Post-training ran in four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning, distillation from six specialist teachers into one model, and agentic reinforcement learning.

For vision, LFM3-VL-2B averages 71.4 across 62 benchmarks, compared with 71.7 for Gemma 4 12B on the same suite. It leads that comparison on visual grounding at 77.3 and GUI understanding at 83.1, while video at 61.2 and agentic tasks at 42.5 trail larger API models. LFM3-Audio-2B reaches 6.2 English word error rate and a 3.8 UTMOSv2 text-to-speech score. LFM3-Omni-2B scores 70.1 on Daily-Omni, against 61.3 for Qwen2.5-Omni-7B. The d1-2B replaces d1-3B with a third fewer parameters and improves Decision Bench from 83.6 to 85.2.

Weights are available on Hugging Face, the Liquid AI Playground and the Liquid API. Documentation lists GGUF, MLX and ONNX for text and vision models, with runtimes including llama.cpp, MLX, vLLM, SGLang, Ollama and Transformers. On-device, Liquid AI measured llama.cpp at Q4_0 with 32K context; a full 128K context needs 1.6 GB of KV cache, compared with 4.3 GB for Qwen3.5-4B. On one NVIDIA B300 with SGLang, LFM3-1.7B produces 3.6 billion tokens per day at 1,024 concurrent requests. Liquid AI also showed specialization: a few hours of supervised fine-tuning and GRPO on one A100 lifted wind-turbine defect classification to 67%.

The comparison in the report lists the LFM Open License v1.0 with a $10 million revenue threshold for LFM3-1.7B, while MiniCPM5-2B, Qwen3.5-4B and Gemma 4 E4B are listed under Apache 2.0.

Editor's Summary

Liquid AI has released LFM3, six open-weight 2B-class models that share a tri-hybrid backbone and cover text, vision, audio, omni and decision tasks. The company reports strong agent and efficiency results for LFM3-1.7B, while larger models still lead on some knowledge and coding benchmarks. The models can be fine-tuned on a single GPU and deployed from phones to cloud.