Liquid AI has released LFM3, a family of six open-weight multimodal models built on a single 2B-class backbone. The suite covers text, vision, audio, omni, and decision tasks, with every variant trainable on a single GPU. Developers can deploy these models on phones, laptops, private servers, or the cloud.
Model specifications and performance
The release includes five specific variants plus two decision models. LFM3-1.7B contains 1.73B parameters and supports a 256K context window. LFM3-VL-2B holds 1.92B parameters with a 128K window. LFM3-Omni-2B reaches 2.2B parameters. The vision model samples video at 2 fps, while the audio model handles speech in and out across 18 languages. The omni model processes vision, audio and language to answer in either text or speech.
Regarding size, the models run on about 1 GB of memory at 4-bit quantisation. GGUF, MLX, ONNX, FP8 and NVFP4 checkpoints are available.
In benchmarks, LFM3-1.7B averages 49.37 across 29 tests. This outperforms Qwen3.5-9B at 47.24 according to Liquid AI’s own evaluation. Specific wins include an IFEval score of 96.66 and an AutomationBench score of 41.29, compared to 15.10 for Qwen3.5-9B. Weaknesses appear in the Humanity’s Last Exam task without tools, scoring 7.17, and the RULER 128K test at 65.56, versus 87.11 for the competitor.
Architecture and training
The backbone mixes three operators across 24 layers. There are 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. The sliding window attends to 256 tokens, while convolution kernels grew from 3 to 7.
Liquid AI used a hardware-in-the-loop architecture search to find this mix. 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 22T tokens. The final 3.5T tokens focused on agent-heavy mid-training, which also extended context to 256K.
Post-training ran in four stages. Supervised fine-tuning on 1.75T tokens came first, then mixed reinforcement learning. Next, six specialist teachers were distilled into one model, followed by agentic reinforcement learning.
Comparison with rivals
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.
Feature comparison
The following table compares LFM3-1.7B against MiniCPM5-2B, Qwen3.5-4B and Gemma 4 E4B:
Parameters: LFM3-1.7B has 1.73B. MiniCPM5-2B has 2.52B. Qwen3.5-4B is a 4B language model (5B on Hub with vision). Gemma 4 E4B is 4.5B effective (8B with embeddings).
Architecture: LFM3 uses Conv + SWA + global attention. MiniCPM5 uses a standard Llama transformer. Qwen3.5 uses Gated DeltaNet + gated attention. Gemma 4 uses Sliding window + global attention.
Context: LFM3 supports 256K. MiniCPM5 supports 131,072. Qwen3.5 supports 262,144. Gemma 4 supports 128K.
Input modalities: LFM3 and MiniCPM5 handle text. Qwen3.5 handles text, image and video. Gemma 4 handles text, image and audio.
License: LFM3 uses the LFM Open License v1.0 with a $10M revenue threshold. The others use Apache 2.0.
Benchmarks: BFCL V4 scores are 71.39 (LFM3), 62.65 (MiniCPM5), 55.12 (Qwen3.5) and 48.25 (Gemma). WebArena-Verified scores are 38.42, 9.11, 21.92 and 17.06 respectively. SWE-bench Verified scores are 43.40, 34.80, 41.20 and 0.50. GPQA Diamond scores are 63.37, 65.15, 75.76 and 57.88. RULER 128K scores are 65.56, 37.96, 84.63 and 18.82. The Liquid 29-benchmark average is 49.37, 39.19, 39.27 and 31.74.
Vision, audio and decision models
LFM3-VL-2B averages 71.4 across 62 vision benchmarks. Gemma 4 12B scores 71.7 on the same suite. It leads that comparison on visual grounding at 77.3 and GUI understanding at 83.1. Video and agentic tasks still trail larger API models, scoring 61.2 and 42.5 respectively.
LFM3-Audio-2B reaches 6.2 English WER and a 3.8 UTMOSv2 TTS score. LFM3-Omni-2B scores 70.1 on Daily-Omni, against 61.3 for Qwen2.5-Omni-7B.
d1-2B replaces d1-3B with a third fewer parameters. It improves Decision Bench from 83.6 to 85.2.
Deployment and commercial terms
Weights are available on Hugging Face, the Liquid AI Playground and the Liquid API. The documentation lists GGUF, MLX and ONNX for the text and vision models. Runtimes include 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. Qwen3.5-4B needs 4.3 GB. On 1 NVIDIA B300 with SGLang, LFM3-1.7B produces 3.6B tokens per day at 1,024 concurrent requests.
Liquid AI also showed specialisation. A few hours of SFT and GRPO on 1 A100 lifted wind-turbine defect classification to 67%.
What it means
Developers can run these models locally without relying on external APIs. The small footprint allows operation on personal devices, while the open weights permit custom fine-tuning for specific tasks like defect detection or agent workflows. The commercial license allows free use up to $10M in annual revenue.
