Liquid AI launches LFM3 family of six open weight class 2B models – Unite.AI

Liquid AI announced LFM3 on October 27, 2026, a new generation of open-weight liquid foundation models that brings text, vision, audio, omni, and decision models onto a single shared backbone. The first release, detailed in a Hugging Face blog post, covers six models in the 2B parameter range.

Six models on a shared backbone

Each LFM3 model branches from the same LFM3-1.7B-Base. LFM3-1.7B accepts text and returns text. The LFM3-VL-2B accepts text, images and video and returns text, while the LFM3-Audio-2B accepts text or audio and responds with text or speech. The LFM3-Omni-2B, which Liquid AI describes as its first omni model, accepts text, images, video and audio, and outputs text and speech. Two decision models, d1-2B and d1-omni-2B, are specialized from the vision and omni models and return calibrated responses rather than generated text.

According to the LFM3-1.7B model sheet, the text model has parameters of 1.73B, a token context length of 256K, a vocabulary of 128,000 tokens, and support for 16 languages, and comes with LFM Open License v1.0. Base and post-trained models are open weight and available on Hugging Face, Liquid AI Playground, and Liquid AI API.

The Liquid AI release post names day one launch partners Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub, and GenieX, and Vercel, and says Mercedes-Benz and Shopify have collaborated for pre-release testing and evaluation.

Tri-hybrid architecture and education

LFM3-1.7B-Base has 24 layers: 13 gated short convolutions, five sliding-window attention layers with a 256-token window, and six global attention layers. Convolutions use a kernel size of 7, compared to 3 in LFM2. Three additional changes define the backbone: head-level controlled attention, in which a sigmoid gate scales the output of each attention head and can suppress activation outliers that impair quantization; partial RoPE, under which the global attention layers rotate only half with the highest frequency of each head while the sliding window layers maintain full RoPE; and a sandwich norm that normalizes the input and output of each block.

The design comes from Liquid AI’s hardware-in-the-loop architecture research, extended for LFM3 to 14 device and backend combinations, including phones (Samsung Galaxy S26 Ultra and OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max and AMD Ryzen AI), Arm and x86 cloud CPUs, Qualcomm Hexagon GPUs and NPUs.

LFM3-1.7B-Base was pre-trained on 22T tokens, with the final 3.5T tokens forming an intermediate training phase on curated data with a larger share of agent traces and tool calls that also extended the context window to 256K tokens. Post-training occurred in four phases: supervised fine-tuning on 1.75T tokens, mixed reinforcement learning, multi-domain policy distillation by six specialized reinforcement learning teachers, and agent-based reinforcement learning within real agent systems.

Liquid AI said that the new architecture allows LFM3-1.7B to outperform its predecessor LFM2.5-2.6B with 35% fewer parameters and that by parameter it matches LFM2-1.2B on short prompts while running about 25% faster with 8K tokens in context. With 4-bit precision, the company said, the text model requires about 1 GB of memory.

Performance and specialization reported by the company

Liquid AI reports that LFM3-1.7B scores 38.42 on WebArena-Verified versus 3.02 for LFM2.5-2.6B, with the largest gains on agent tasks, although the model sheet states that the model is not recommended for agent coding or complex multi-step reasoning due to its small size.

Liquid AI said the LFM3-VL-2B is the first LFM vision model to accept video input and reports that it increases the average across 62 vision benchmarks from 57.6 to 71.4 compared to the LFM2.5-VL-3B using parameters of 1.9B versus 3.1B, putting it on par with the Gemma 4 12B, a model more than six times larger. A new internal 164M parameter vision encoder uses 3x less processing than the SigLIP2 encoder in previous LFM VL models and matches the accuracy of the Qwen3.5-2B’s vision encoder with 2.2x fewer FLOPs, according to the company’s measurements.

LFM3-Audio-2B supports 18 languages, and Liquid AI reports speech quality of 3.83 on UTMOSv2 versus 3.22 for LFM2.5-Audio-1.5B, a fully random audio encoder with time to first token as low as 50 ms, and a LiquidCodec tokenizer that generates speech at approximately 87 times in real time on a single CPU thread on an Apple M4 Max. The company said that the parameter The 2.2B LFM3-Omni-2B beats models with parameters up to 10.7B, including the Qwen2.5-Omni-7B, on the Daily-Omni, WorldSense, and AV-SpeakerBench audiovisual comprehension benchmarks.

Decision models answer a list of questions in a single forward pass, each with a calibrated probability, and add typed responses that point to objects, read values ​​from documents, estimate numbers with a range, and choose each applicable option. Liquid AI said d1-2B replaces d1-3B with a third fewer parameters and scores 75.5 versus 64.0 on ImajevBench.

Each LFM3 model can be tuned on a single GPU, and the notebooks published by Liquid AI cover supervised tuning, DPO, GRPO, and continuous pre-training for text, vision, audio, and omni models. In a company example, LFM3-VL-2B was fine-tuned on approximately 3,000 public photos of wind turbine blades over a few hours on an A100, increasing defect classification accuracy to 67% and F1 detection from 0.003 to 0.11 on photos held outside.

Liquid AI also adapted the LFM3-VL-2B into three robotics models for orchestration, navigation and manipulation. The company reports that the vision-language-action manipulation model, trained in 15 embodiments, achieved an average success of 96.2% in the four LIBERO suites and 45.7% in the RoboCasa-GR1 tabletop tasks while generating action blocks in less than 50 ms on a Jetson AGX Orin with TensorRT FP16.

Regarding throughput, Liquid AI said the LFM3-1.7B can produce 3.6 billion tokens per day on a single NVIDIA B300 at 1,024 concurrent requests. Models ship with native support for llama.cpp, MLX, and vLLM on Apple, AMD, NVIDIA, and Qualcomm hardware, with checkpoint variants including BF16, FP8, NVFP4, GGUF, MLX, and ONNX.