Liquid AI released six small LFM3 models

Liquid AI announced its new model family, called LFM3, on October 27, 2026. All six models are derived from the same base, LFM3-1.7B-Base. The company’s goal is to process text, images, video, and audio in the same architecture and run them on devices such as phones, laptops, and single-board computers.

The models are open-weight and available under the LFM Open License v1.0 via Hugging Face, the Liquid AI Playground, and the company’s API.

One base, six different models

The family’s lineup is function-focused. LFM3-1.7B takes text and returns text. LFM3-VL-2B accepts text, images, and video. LFM3-Audio-2B takes text or audio and produces text or speech. LFM3-Omni-2B is the company’s first multimodal (omni) model: it processes text, images, video, and audio together and outputs text and speech.

That leaves two decision models. d1-2B and d1-omni-2B respond with calibrated probabilities instead of generated text. These models are designed for tasks such as reading a value from a document, object detection, estimating a number as a range, or marking all valid options.

Triple hybrid architecture

LFM3-1.7B-Base consists of 24 layers: 13 gated short convolution layers, five sliding-window attention layers with a 256-token window, and six global attention layers. The convolution kernel size was increased from 3 in LFM2 to 7.

The architecture has three distinguishing features. Head-based gated attention can scale each attention head’s output with a sigmoid gate, suppressing outliers that degrade quantization quality. In partial RoPE, the global attention layers rotate only the high-frequency half of each head, while the sliding-window layers continue to use full RoPE. Sandwich normalization, meanwhile, normalizes each block’s input and output separately.

This design was derived through hardware-in-the-loop architecture search. For LFM3, Liquid AI expanded the search to 14 device and backend combinations: phones such as the Samsung Galaxy S26 Ultra and OnePlus 15, the Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, Arm and x86 cloud processors, GPUs, and the Qualcomm Hexagon NPU.

Training and reported results

The base model underwent pretraining on 22 trillion tokens. The final 3.5 trillion tokens were used as an intermediate training stage weighted toward agent traces and tool calls; the context window was also extended to 256,000 tokens during this stage. This was followed by supervised fine-tuning, mixed reinforcement learning, multi-domain distillation from six expert teachers, and reinforcement learning in real agent environments.

According to the company’s measurements, the text model, with 1.73 billion parameters, performs better than the 2.6-billion-parameter LFM2.5, using 35% fewer parameters. At an 8,000-token context, it runs about 25% faster than the previous model and uses about 1 GB of memory at 4-bit quantization. The model card lists support for 16 languages.

On the WebArena-Verified test, LFM3-1.7B scores 38.42 while the older model stays at 3.02. However, the model card explicitly states that, because of its small size, the model is not recommended for agentic coding and complex multi-step reasoning. The vision model, LFM3-VL-2B, raises the average across 62 vision tests from 57.6 to 71.4 and is claimed to match Gemma 4 12B, which is six times larger. In the audio model, the speech quality metric UTMOSv2 score rises to 3.83.

Edge deployment and ecosystem

The practical side of the family is that it can be fine-tuned on a single GPU. Liquid AI published notebooks for supervised fine-tuning, DPO, GRPO, and additional pretraining. In the company’s example, the vision model was trained for a few hours on a single A100 using about 3,000 wind turbine photos, yielding a significant improvement in defect detection.

The models are supported locally with llama.cpp, MLX, and vLLM; there are different checkpoint formats such as BF16, FP8, NVFP4, GGUF, MLX, and ONNX. Launch partners include AMD, Arm, Intel, NVIDIA, Qualcomm, Ambarella, OpenRouter, and Vercel. Mercedes-Benz and Shopify participated in pre-release testing.

Liquid AI’s direction is clear: instead of competing with giant models, it aims to produce small models that run on constrained hardware and do specific tasks well. This approach makes sense especially in scenarios where data never leaves the device and in applications where cloud costs are a problem. However, all performance figures are based on the company’s own measurements; independent benchmarks will show where the model really stands.

Source: Unite.AI