By Hamza Faisal
Liquid AI has released LFM3, a family of six open-weight artificial intelligence models built around two billion parameters, designed to run agents and multimodal tasks on phones, laptops and single-board computers rather than only in the cloud. The MIT spinout announced the family on October 27, pitching it as a suite for specialised intelligence: pick the model that fits the input and the device, fine-tune it on your own data, and run it locally.
All six models share a single backbone, LFM3-1.7B-Base, and a new tri-hybrid architecture. Inside its 24 layers, the design mixes 13 gated short-convolution layers, five sliding-window attention layers and six global attention layers, with a 256,000-token context window for the text model. Liquid AI says the architecture was optimised across 14 processor classes, from Arm and x86 CPUs and Qualcomm Hexagon neural processors to GPUs, covering devices including the Galaxy S26 Ultra, the Raspberry Pi 5 and laptops running Apple M4 Max and AMD Ryzen AI chips.
Six models, one backbone
The headline model, LFM3-1.7B, carries 1.73 billion parameters and fits in about one gigabyte of memory at 4-bit precision, small enough to run local agents on personal hardware without per-token API fees. LFM3-VL-2B adds a lightweight vision encoder for images and video, with step-by-step reasoning and video input at two frames per second across a 128,000-token context. LFM3-Audio-2B handles spoken requests and tool calls in 18 languages, while LFM3-Omni-2B, the company's first omni model, combines text, images, video and audio with speech output in a 2.17-billion-parameter package.
The remaining pair takes a different approach. The d1-2B and d1-Omni-2B decision models answer several questions in a single forward pass and return structured results with calibrated probabilities: yes-or-no answers, multiple-choice picks, bounding boxes on images, dates and estimated ranges, aimed at factory and business workflows rather than open-ended chat.
Punching above their weight
On Liquid AI's own benchmarks, the small models compare surprisingly well with much larger ones. The text model averages 49.37 across text benchmarks, ahead of Qwen3.5-9B at 47.24, and scores 38.42 on the WebArena-Verified agentic benchmark, against 3.02 for its predecessor LFM2.5-2.6B. The vision model averages 71.4 across 62 benchmarks, matching Google's Gemma 4 12B, a model six times its size, while the omni model beats Qwen2.5-Omni-7B on the Daily Omni benchmark. The company also reports faster on-device decoding than rivals and a smaller memory footprint: 1.6 gigabytes of KV cache at a 128,000-token context, against 4.3 gigabytes for Qwen3.5-4B.
Liquid AI is candid about the limits. The 2B-class models trail larger systems on deep trivia tests such as GPQA Diamond and on complex coding and mathematics, and the company recommends pairing them with search tools rather than using them for heavy reasoning alone.
Fine-tune on one GPU, use commercially
For developers, the practical pitch is accessibility. Every model can be fine-tuned on a single GPU, and the family works with llama.cpp, MLX, vLLM, SGLang, ONNX, Ollama and LM Studio, available on Hugging Face, Liquid AI's Playground and its API. The models are released under the LFM Open License v1.0, which allows free commercial use for companies with under $10 million in annual revenue.
For larger companies, that trade-off is exactly the point: as frontier models converge on capability, the contest is shifting to cost, latency and where the model runs, and LFM3 is a bet that the next wave of AI agents will live on the device, not only in the data centre.



