Liquid AI packs text, vision, audio and decision models into one 2B backbone, about 25% faster than LFM2 at 8K context
Synopsis
Liquid AI released LFM3, a family that puts text, vision, audio, omni and decision models on one shared backbone of roughly 2B parameters, with base and post-trained weights open-weight on Hugging Face and its API. The backbone mixes gated short convolutions, sliding-window attention and global attention, was pre-trained on 22T tokens, and a final 3.5T-token mid-training stage extends context to 256K. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.
Interpretation
LFM3 places text, vision, audio, omni and two decision models on one shared backbone of about 2B parameters; the first release has six models, all branched from LFM3-1.7B-Base. LFM2's hybrid architecture was built for fast on-device inference, while LFM3 redesigns it for agentic use, which needs more capability per parameter and fast inference over long contexts. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.
The new backbone combines three operators: of 24 layers, 13 are gated short convolutions, 5 are sliding-window attention with a 256-token window, and 6 are global attention, with convolution kernel size raised from 3 in LFM2 to 7. LFM2 had shown that a hybrid of short convolutions and attention could run fast on-device without sacrificing quality; LFM3 swapped some sequence blocks and found sliding-window attention gave the best quality-speed tradeoff. The hardware-in-the-loop search expanded from CPUs in one phone and one laptop for LFM2 to 14 device and backend combinations for LFM3, including phones, a Raspberry Pi 5, laptops, Arm and x86 cloud CPUs, GPUs and the Qualcomm Hexagon NPU.
The text model LFM3-1.7B targets local agents with a 256K-token context window and is post-trained to run in agentic harnesses such as Hermes Agent, Pi and OpenClaw. Measured on-device with llama.cpp at Q4_0 and a 32K-token context, LFM3-1.7B is the fastest model on all three devices, decoding 1.8 to 2.8 times faster than the next one, and uses the least memory. A full 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B; on a single NVIDIA B300 with SGLang in BF16 it can produce 3.6 billion tokens per day at 1,024 concurrent requests.
Perspective
The demonstrated scope is the first six roughly 2B-parameter models and their measurements on edge devices, a single B300, and simulated and real-robot data; deeper performance and inference results are deferred to the individual model cards.
The release states that larger models still win on agentic coding and knowledge, where size matters most, and that search only narrows the knowledge gap, nearly doubling the Humanity's Last Exam score; the vision model remains within about 2 points of GPT-5.6 Luna.