LFM3: Liquid AI releases six small open models of 2 billion parameters, from text to audio, to run on-device
Liquid AI launches LFM3, a family of models of roughly 2 billion parameters (text, vision, audio, omni and decision) with downloadable weights, designed for local agents. Watch the license: commercial use is reserved for companies with less than $10 million in revenue.
Liquid AI, which designs compact models meant to run on-device, launched this Tuesday, October 27, LFM3, its third generation of Liquid Foundation Models. The official announcement dropped on X at 6:07 PM, Paris time, alongside a lengthy blog post. The weights are downloadable starting today on Hugging Face, and the models are also available in the Playground and the company’s API.
The bet runs counter to the race for very large models: LFM3 brings together six models of roughly 2 billion parameters, each specialized, designed to run directly on a phone, laptop or embedded board, without going through the cloud.
One family, six uses
All the models are derived from the same base, LFM3-1.7B-Base, also released for those who want to train it their own way:
- LFM3-1.7B: the text model, with a 256,000-token context window, post-trained to drive agents (tools, web browsing, office tasks);
- LFM3-VL-2B: vision and language, which accepts video input for the first time at Liquid AI;
- LFM3-Audio-2B: voice conversation, transcription, synthesis and translation in 18 languages, with tool calling;
- LFM3-Omni-2B (2.2 billion parameters): the company’s first “omni” model, combining image, sound and text;
- d1-2B and d1-Omni-2B: decision models that, instead of writing out a response, answer a list of questions in a single pass (yes/no, multiple choice, field extraction, object localization), each with a probability.
What’s changing under the hood
Liquid AI keeps its hybrid recipe but enriches it. Across the text model’s 24 layers, 13 are short convolutions, 5 use sliding-window attention of 256 tokens and 6 use global attention. The stated goal is to stay fast on long contexts: according to the developer, a full 128,000-token context takes up only 1.6 GB of memory cache, versus 4.3 GB for Qwen3.5-4B. Pre-training covers 22 trillion tokens, with more code, math and agent traces than the previous generation.
The architecture has been optimized on 14 hardware combinations, from Samsung and OnePlus smartphones to the Raspberry Pi 5, including Apple, AMD and Qualcomm chips and GPUs. The text model fits in about 1 GB at 4-bit, and each model can be fine-tuned on a single GPU. On the tooling side, llama.cpp, MLX and vLLM are supported from launch, with partners including AMD, Qualcomm, NVIDIA, Arm, OpenRouter and Vercel.
Flattering scores, to be verified
The published numbers are Liquid AI’s, with no independent measurement for now. On its own table, LFM3-1.7B scores an average of 49.37, ahead of Qwen3.5-9B (47.24), a model more than five times larger. The gap is clear on browser agent tasks (38.42 versus 28.94 on WebArena-Verified).
The table also shows its limits, and the developer acknowledges it: on agentic coding (43.40 versus 51.20 on SWE-bench Verified) and knowledge (63.37 versus 80.40 on GPQA Diamond), larger models keep the edge. For vision, LFM3-VL-2B would show an average of 71.4 on 62 tests, on par with Gemma 4 12B (71.7), a model six times larger.
The point not to miss: the license
Liquid AI talks about “open” models, but these are downloadable weights under the LFM Open License v1.0, not a free license like Apache 2.0 or MIT. The text, available on Hugging Face, reserves commercial use for entities with annual revenue below $10 million. Beyond that, you need to reach an agreement with Liquid AI. For an independent developer or a startup, there’s no immediate consequence; for a large company, it’s a blocking point to check before any deployment.
Our take
LFM3 illustrates a deep trend: rather than one large generalist model, a series of small specialized models, light enough to run locally and easy enough to fine-tune to become experts at a specific task. If the scores hold up, LFM3-1.7B would be one of the best choices for running agents on a laptop without paying for an API. But two caveats are worth keeping in mind: the comparisons all come from the developer, and the license is not free for large companies. The first independent tests will tell whether the claimed gain on agent tasks holds outside Liquid AI’s benchmarks.
Sources
- Liquid AI — Introducing LFM3: Every Modality for Specialized Intelligence Primary source
- Liquid AI (X) — LFM3 announcement Primary source
- Hugging Face — LiquidAI/LFM3-1.7B model card Primary source
- Hugging Face — LFM3 collection Primary source
- Hugging Face — LFM Open License v1.0 (LFM3-1.7B) Primary source — Commercial use limitation clause ($10 million threshold)
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
