Liquid AI Releases LFM3, Six Small Open-Weight Models for Any Device

Liquid AI has released LFM3, a family of six small AI models that handle text, images, video, speech and structured decisions, and which the company says can run on phones and laptops and be retrained on a single graphics card.

The company announced the models on 27 October in a blog post, calling LFM3 “our new generation of Liquid Foundation Models (LFMs) built for specialized intelligence”. All six are open-weight, meaning the trained model files can be downloaded, and they are available on Hugging Face and through Liquid AI’s own API. Every performance figure in this story comes from Liquid AI’s own tests, which have not been independently checked.

Six models on one base

The first LFM3 release covers six models, all branched from the same 1.7-billion-parameter base model, according to Liquid AI’s Hugging Face post:

• LFM3-1.7B, a text model aimed at agents, tool calling and structured output.
• LFM3-VL-2B, a vision model and the first in the series to take video as input.
• LFM3-Audio-2B, which listens and speaks, and handles speech recognition, speech synthesis and translation across 18 languages.
• LFM3-Omni-2B, the company’s first “omni” model, which takes text, images and audio and replies in text or speech.
• d1-2B and d1-omni-2B, “decision” models that answer a list of questions about a situation, such as a support ticket or a photo, “in a single forward pass, each with a calibrated probability”. TSN covered the first d1 models earlier this month.

“Instead of running one large general-purpose model, you can pick the one that fits your use case, customize it, and run it on-device,” the company writes.

A new design for long tasks

Liquid AI says LFM3 uses a new “tri-hybrid” design. Of the base model’s 24 layers, 13 are gated short convolutions, a cheap way of looking at nearby words; 5 use attention over a sliding window of 256 tokens; and 6 use full attention across the whole input. Attention is the part of a language model that relates each word to others, and it gets expensive as inputs grow.

The base model was pre-trained on 22 trillion tokens of text, the company says, with the final 3.5 trillion focused on examples of agents using tools. That stage also stretched the context window, the amount of text the model can consider at once, to 256,000 tokens.

According to Liquid AI, the text model fits in about 1 GB of memory when compressed to 4-bit precision, and with 8,000 tokens in context it runs about 25% faster per parameter than LFM2-1.2B. The company says it tuned the design by testing it on 14 combinations of hardware and software, including Samsung and OnePlus phones, a Raspberry Pi 5, Apple and AMD laptops and Qualcomm’s Hexagon NPU chip.

What Liquid AI claims on performance

The company’s headline claim is that LFM3-1.7B outperforms its predecessor, LFM2.5-2.6B, “while using 35% fewer parameters”. On WebArena-Verified, a test of agents completing tasks in websites, it reports a score of 38.42 against 3.02, and says the model “even beats Qwen3.5-9B on long tasks in browsers and business apps”.

Liquid AI’s model card backs that up on its chosen tests. It shows LFM3-1.7B at 38.42 on WebArena-Verified against 28.94 for Qwen3.5-9B, and 41.29 against 15.10 on AutomationBench. The same table shows the larger Qwen model still ahead on others, including MCP-Atlas, a tool-use test (42.67 to 32.57), and SWE-bench Verified, a coding test (51.20 to 43.40).

For vision, the company says LFM3-VL-2B raises the average across 62 benchmarks “from 57.6 to 71.4” compared with its 3.1-billion-parameter predecessor, and puts it level with Google’s Gemma 4 12B, a model more than six times larger.

Who is lining up behind it

Liquid AI lists AMD, Arm, Intel, NVIDIA, Qualcomm, Ambarella, OpenRouter and Vercel as launch partners, and says Mercedes-Benz and Shopify took part in testing before release.

“As AI increasingly moves from the cloud to the client, developers need models that can be deployed efficiently across the broad spectrum of AI PCs,” said Josh Hort, a senior director at AMD, in a statement published by Liquid AI. Jörg Burzer, Mercedes-Benz’s chief technology officer, said the carmaker is “exploring how powerful on-device AI can further enhance the MBUX Virtual Assistant”, its in-car voice assistant. Neither statement describes a product commitment.

Why Liquid AI is betting on small specialists

LFM3 is a bet against the idea that one giant model should do everything. “Frontier models are powerful generalists, but they are not necessarily experts in your use case,” Liquid AI argues, and its pitch is that a small model, retrained on one graphics card with a company’s own data, can become a dependable expert that runs where the data already sits. That matters for phones, cars and factory equipment, where sending every request to a data centre adds delay, cost and privacy risk.

The release also continues a clear pattern. Two weeks ago Liquid AI shipped LFM2.5-24B-A2B, a larger model built for agents on laptops. Others are chasing the same space, from Recursive’s locally run Flint coding model to chip designers such as BrainChip, which is offering designs for running language models on battery-powered devices.

Two caveats apply. The benchmark numbers are Liquid AI’s own, on tests it chose, and its own table shows a model more than five times larger still winning on coding and some tool use. And “open-weight” is not the same as open source: under the LFM Open License, commercial use is licensed only for organisations below $10 million in annual revenue, so larger companies that want to build products on LFM3 will need a separate deal with Liquid AI.

Sources

1. Liquid AI, “Introducing LFM3: Every Modality for Specialized Intelligence”, 27 October 2026. https://www.liquid.ai/blog/lfm3
2. Liquid AI on Hugging Face, “Build specialized multimodal models with LFM3”, 27 October 2026. https://huggingface.co/blog/LiquidAI/lfm3
3. Liquid AI, LFM3-1.7B model card, Hugging Face, 27 October 2026. https://huggingface.co/LiquidAI/LFM3-1.7B
4. Liquid AI, LFM Open License v1.0, Hugging Face. https://huggingface.co/LiquidAI/LFM3-1.7B/blob/main/LICENSE
5. Liquid AI (@liquidai) on X, 27 October 2026. https://x.com/liquidai/status/2115128329003303107

Related stories

• Liquid AI Releases LFM2.5-24B-A2B, an Open-Weight Model for Running AI Agents on Your Own Device
• Small Open Models That Answer in One Pass: Liquid AI’s d1 and Perplexity’s pplx-embed-v2-late
• Recursive Releases Flint, an Open-Weight Coding Model Built to Run Locally
• BrainChip Launches Sequora, Chip Designs for Running Language Models on Battery-Powered Devices