Liquid AI has unveiled LFM3, a new generation of open-weight Liquid Foundation Models built around one shared backbone. Released on October 27, 2026, the family brings text, vision, audio, omni, and decision capabilities into a six-model lineup in the 2B parameter range.
That shared foundation is the story. Every LFM3 model branches from LFM3-1.7B-Base, giving the family a common architectural core while allowing each model to handle a different kind of input, output, or task.
Six Models, One Common Foundation
The first LFM3 release covers models designed for text, images, video, audio, speech, and calibrated answers. LFM3-1.7B takes text in and returns text, while LFM3-VL-2B accepts text, images, and video and returns text.
LFM3-Audio-2B takes text or audio and responds with text or speech. LFM3-Omni-2B expands that multimodal reach further, accepting text, images, video, and audio while producing text and speech. It is Liquid AI’s first omni model.
The lineup also includes two decision models: d1-2B and d1-omni-2B. Liquid AI specialized d1-2B from the vision model and d1-omni-2B from the omni model, with both designed to return calibrated answers rather than generated text.
According to the LFM3-1.7B model card, the text model has 1.73B parameters, a 256K-token context length, a 128,000-token vocabulary, and support for 16 languages. It ships under the LFM Open License v1.0.
The base and post-trained models are open-weight and available on Hugging Face, the Liquid AI Playground, and the Liquid AI API. That gives the same family a path across model access, interactive use, and API deployment.
An Architecture Built Around Real Hardware
LFM3-1.7B-Base contains 24 layers, including 13 gated short convolutions, five sliding-window attention layers with a 256-token window, and six global attention layers. The convolutions use a kernel size of 7, up from 3 in LFM2.
The design also adds head-wise gated attention, which scales each attention head’s output with a sigmoid gate. That mechanism can suppress activation outliers that hurt quantization, a key detail for models intended to run across different devices and backends.
Partial RoPE forms another part of the design. It rotates only the highest-frequency half of each head in global attention layers, while sliding-window layers keep full RoPE. A sandwich norm normalizes each block’s input and output.
Liquid AI developed the design through hardware-in-the-loop architecture search, expanding the search to 14 device and backend combinations for LFM3. Those combinations included phones such as Samsung Galaxy S26 Ultra and OnePlus 15, a Raspberry Pi 5, and laptops with Apple M4 Max and AMD Ryzen AI.
The search also covered Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU. The result is an architecture shaped around a broad hardware map rather than a single deployment target.
Training Scale Meets Smaller Model Efficiency
LFM3-1.7B-Base was pre-trained on 22T tokens. The final 3.5T tokens formed a mid-training stage on curated data, with a larger share of agentic traces and tool calls, while that stage extended the context window to 256K tokens.
Post-training ran in four stages. The process included supervised fine-tuning on 1.75T tokens, mixed reinforcement learning, multi-domain on-policy distillation from six specialist reinforcement-learning teachers, and agentic reinforcement learning inside real agent harnesses.
Liquid AI said the new architecture lets LFM3-1.7B outperform its predecessor, LFM2.5-2.6B, with 35% fewer parameters. The company also said it matches LFM2-1.2B per parameter on short prompts and runs about 25% faster with 8K tokens in context.
That combination points to the role Liquid AI wants LFM3 to play: one model family that can move from language tasks to multimodal interaction and calibrated decision-making, while staying grounded in hardware-aware design.
Day-0 launch partners include Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub and GenieX, and Vercel. Mercedes-Benz and Shopify partnered on pre-release testing and evaluation.
LFM3 now puts six open-weight models behind one shared backbone, with access through Hugging Face, the Liquid AI Playground, and the Liquid AI API. The next phase will show how this common foundation performs across phones, laptops, edge devices, cloud systems, and the multimodal applications built around it.
Based on
- Liquid AI Launches LFM3 Family of Six Open-Weight 2B-Class Models — unite.ai
- Riemann Computing Inc. Creates and Demos Out a 13B Parameter AI Model, Aims at Massively Reducing Training Costs | AP News — apnews.com
- Hugging Face’s New AI Model Carbon-X Analyzes, Generates DNA Sequences – Business Insider — businessinsider.com
- Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox | AP News — apnews.com
- From Words To Actions: How LLMs Are Reshaping Robotics — forbes.com
