2B Models Take On 12B: Liquid AI Releases the Full LFM3 Family

Liquid AI has released LFM3, a family of open-weight 2B-class models. The six model instances cover text, vision, audio, omni-modal, and structured decision-making. The official positioning is specialized intelligence: choose a model for the scenario, fine-tune it, and deploy it on-device.

LFM3 uses a new tri-hybrid architecture. Rather than selecting structures using only public benchmarks, Liquid AI used its in-house hardware-in-the-loop architecture search, putting inference speed, memory footprint, and model quality together into the search objective. The search covered 14 device configurations: phones, laptops, Raspberry Pi 5, cloud CPUs, GPUs, and Qualcomm Hexagon NPUs. Deployment constraints were factored in when choosing the architecture.

Text Model LFM3-1.7B

It has a 256K context, supports 16 languages, and was post-trained for agent harnesses. At 128K tokens, the KV cache (the key-value cache used during inference) takes up 1.6 GB, compared with 4.3 GB for Qwen3.5-4B. It can be fine-tuned on a single GPU and is also suitable for running local agents: inbox classification, calendar management, web research, and tool calling. Officially, it was used with 300,000 parallel agents to analyze the Enron email archive. On WebArena-Verified and AutomationBench, the 1.7B model surpasses Qwen3.5-9B. It can also handle server-side workloads; the company says p95 latency remains stable under high pressure, making it suitable for latency-sensitive scenarios such as online fraud detection. Liquid AI also notes that it ranks first on the Pareto frontier for device intelligence.

Vision Model LFM3-VL-2B

1.9B parameters, supports image and video input, and can perform step-by-step reasoning, chart math, and spatial grounding. It averages 71.4 across 62 vision benchmarks, on par with Gemma 4 12B. The in-house vision encoder uses roughly one-third of the compute of SigLIP2. The company has drawn boundaries: it is not suitable for coding or complex multi-step reasoning.

Audio Model LFM3-Audio-2B

Supports 18 languages; accepts text or audio and outputs text, speech, or mixed content. Its causal audio encoder can process and prefill while the user is speaking, and the company says first-token latency can be as low as 50 ms. LiquidCodec synthesizes speech at about 87× real time on a single CPU thread of an Apple M4 Max. On an internal spoken tool-calling evaluation, it scores 58.5, compared with 44.3 for Qwen2.5-Omni-3B.

Omni-Modal Model LFM3-Omni-2B

2.2B parameters, the first omni model to bring vision, audio, and language together. It can see images and hear audio, then output text or speech. It scores 70.1 on Daily-Omni, compared with 61.3 for Qwen2.5-Omni-7B.

Decision Models d1-2B and d1-Omni-2B

The former accepts text and images, while the latter adds audio. Both convert inputs into structured answers, producing results with calibrated probabilities in a single forward pass. They support bounding boxes, text/number/date extraction, returning not found, and numeric range estimation. On 117 benchmarks, d1-2B scores 63.2, even higher than d1-3B. Some developers think this is best suited for classification and routing: a single threshold decision is all it takes. Other developers caution that fine-tuning the d1 series may break calibrated probabilities; it is best to refit temperature on a held-out set and then check reliability diagrams.

Fine-Tuning Results

A concrete example: using about 3,000 public WTBs2025 wind turbine blade photos, they performed SFT and GRPO on LFM3-VL-2B, training for a few hours on a single A100. On about 1,400 held-out photos, defect classification accuracy reached 67%, while both the original model and GPT-6 Luna were below 10%. Detection F1 rose from 0.003 to 0.11. The model can also write inspection conclusions, describing the defect, evidence, and location.

Robotics is being explored as well: LFM3-VL-2B handles visual navigation and manipulation, paired with d1-2B for skill monitoring. A VLA (vision-language-action model) trained across 15 embodiments averages 96.2% on the four LIBERO suites and 45.7% on RoboCasa-GR1 tabletop tasks, and generates action chunks in under 50 ms on a Jetson AGX Orin using TensorRT FP16.

LFM3 is already available on Hugging Face, with BF16, FP8, NVFP4, GGUF, MLX, ONNX, and other versions, and supports vLLM, SGLang, llama.cpp, Ollama, and LM Studio. AMD, Arm, NVIDIA, OpenRouter, Qualcomm, Intel, and others are day-0 partners.

In the comments, developers are already planning to fine-tune one over the weekend and launch it on Sunday. Some are asking about memory usage after quantization, and others say 2B is the sweet spot for fine-tuning. Whether a model is small or not doesn't matter; where it can run and whether it can be shaped to fit your needs—that's what counts.