Liquid AI Releases Open Weight LFM3 Family of Six 2B Multimodal and Decision Models Built on a Tri Hybrid Architecture
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence. > 6 model instances, across text, audio, vision, and omni modalities, generative and decision models. > 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window. > 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding > 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages. > 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language. > 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input. > fine-tune on 1 GPU. > a new tri-hybrid architecture > architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs. > 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters. > 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒 Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device. 1/ 🧵
Context
Packing text, video vision, streaming speech, cross modal omni understanding, and single pass calibrated decision making into 2B parameter open weight models that fit in 1 gigabyte at 4 bit precision brings capable multimodal agents directly onto phones, laptops, and embedded robotics hardware.
Liquid AI previously demonstrated with its LFM2 and LFM2.5 releases that hybrid architectures combining short convolutions and attention can run efficiently on edge CPUs without sacrificing instruction following quality. In August 2026, the company partnered with Artificial Analysis to launch Pipette, an open source benchmarking platform measuring on device model throughput, latency, and memory across consumer phones and laptops.
Deep dive
On October 27, Liquid AI released LFM3, a suite of six open weight 2B class foundation models built on a single shared backbone, LFM3 1.7B Base. Designed via hardware in the loop architecture search across 14 processor and backend combinations including Galaxy S26 Ultra, Apple M4 Max, AMD Ryzen AI, Raspberry Pi 5, and Qualcomm Hexagon NPUs, the 24 layer backbone introduces a tri hybrid mix of 13 gated short convolutions with kernel size 7, 5 sliding window attention layers, and 6 global attention layers alongside head wise gated attention, partial RoPE, and sandwich normalization.
Pretrained on 22 trillion tokens with a 3.5 trillion token agentic mid training stage, LFM3 1.7B supports a 256K context window, fits in roughly 1 gigabyte at 4 bit quantization, and scores 38.42 on WebArena Verified versus 3.02 for LFM2.5 2.6B while beating Qwen3.5 9B on browser and business automation tasks. At 128K context, its key value cache consumes 1.6 gigabytes compared with 4.3 gigabytes for Qwen3.5 4B.
Branching from the same base, the 1.9B parameter LFM3 VL 2B adds video input and a 164M parameter vision encoder that uses one third the compute of SigLIP2, averaging 71.4 across 62 vision benchmarks to match Gemma 4 12B at one sixth the size. LFM3 Audio 2B pairs a causal audio encoder with LiquidCodec, decoding speech at 87 times real time on a single Apple M4 Max CPU thread across 18 languages. LFM3 Omni 2B unifies vision, audio, and language at 2.2B parameters, scoring 70.1 on Daily Omni versus 61.3 for Qwen2.5 Omni 7B. Rounding out the suite, d1 2B and d1 Omni 2B act as single pass decision models returning calibrated probabilities, bounding boxes, and extracted spans.
What to expect next
Because every LFM3 checkpoint ships open weight on Hugging Face with day zero support across llama.cpp, vLLM, SGLang, MLX, Ollama, and Unsloth and fine tunes on a single GPU, edge developers can immediately replace multi model voice and vision pipelines with unified 2B checkpoints. Watch for robotics builders to adopt the LFM3 VL 2B manipulation adaptation, which generates action chunks in under 50 milliseconds on Nvidia Jetson AGX Orin.