Liquid AI releases LFM3: six multimodal 2B models with a 256K context window

Liquid AI has released LFM3, a model generation built around one idea: specialized intelligence should live next to the application, not in the cloud. Six 2B-class models cover text, vision, audio, omni mode and calibrated answers. All of them branch from a single base, LFM3-1.7B-Base, and each can be fine-tuned on one GPU. The text model LFM3-1.7B beats the previous LFM2.5-2.6B with 35% fewer parameters, takes up about 1 GB with 4-bit quantization and holds a 256K-token context.

Three layer types instead of two

LFM2 showed that a hybrid of short convolutions and attention runs fast on-device without losing quality. For LFM3, the team again ran an architecture search with hardware-in-the-loop feedback, this time across 14 device and backend combinations: from phones and Raspberry Pi 5 to laptops, cloud CPUs, GPUs and Qualcomm Hexagon NPUs. The goal was to give agents more capability per parameter and fast inference at long context.

LFM3-1.7B-Base has 24 layers, arranged as follows:

• 13 gated short convolutions with kernel size 7 instead of the 3 used in LFM2;
• 5 sliding-window attention layers with a 256-token window;
• 6 global attention layers;
• head-level attention gating: a sigmoid scales the output of each head, so a head can switch itself off when it has nothing to add. This helps suppress activation outliers that interfere with quantization;
• partial RoPE in the global layers: only the high-frequency half of each head is rotated, while the other half matches content regardless of distance;
• sandwich norm: normalization at both the input and the output of each block.

The payoff shows up at length. On short prompts, LFM3-1.7B matches LFM2-1.2B on-device per parameter, and at 8K tokens of context it runs about 25% faster. Pretraining took 22T tokens with a 128K vocabulary, the same as LFM2.5-8B-A1B. The last 3.5T tokens went to mid-training on curated data with a higher share of agent traces and tool calls, and the context was extended to 256K at the same stage.

Six models and two new encoders

Instead of one large general-purpose model, Liquid lets you pick the one that fits your input data:

• LFM3-1.7B: text in and out, built on LFM3-1.7B-Base;
• LFM3-VL-2B: text, images and video in, text out; Liquid's first vision model with video;
• LFM3-Audio-2B: text and audio in, text and speech out; 18 languages;
• LFM3-Omni-2B: text, images, video and audio in, text and speech out; the company's first omni model;
• d1-2B: built on LFM3-VL-2B, produces calibrated answers;
• d1-omni-2B: the same on top of LFM3-Omni-2B, with audio input.

The 164M-parameter vision encoder combines convolutions and attention and needs three times less compute than the SigLIP2 used in earlier VL models. Its accuracy matches the Qwen3.5-2B encoder, but it uses 2.2 times fewer FLOPs (measured on the same LFM2.5-230M language backbone). For video, a pair of selected frames is collapsed into a single temporal slice, which halves the sequence length.

The audio encoder mixes sliding-window attention with short convolutions and is fully causal, so the model fills the prefill while the user is still speaking. Time to first token starts at 50 ms. The new LiquidCodec tokenizer turns tokens into speech about 87 times faster than real time on a single thread of an Apple M4 Max CPU, and it supports streaming decoding.

The numbers that matter for agents

On WebArena-Verified, LFM3-1.7B scores 38.42 vs 3.02 for LFM2.5-2.6B. On long tasks in the browser and in business applications, it beats Qwen3.5-9B. The model is fine-tuned for the Hermes Agent, Pi and OpenClaw agent harnesses.

LFM3-VL-2B (1.9B parameters vs 3.1B for LFM2.5-VL-3B) lifts the average across 62 vision benchmarks from 57.6 to 71.4 and matches Gemma 4 12B, which is more than six times larger. The model reasons step by step, locates objects and estimates their relative positions, which robotics needs. LFM3-Audio-2B improves synthesis quality from 3.22 to 3.83 on UTMOSv2 relative to LFM2.5-Audio-1.5B, and adds audio understanding and tool calling from voice requests. On Daily-Omni, WorldSense and AV-SpeakerBench, LFM3-Omni-2B beats models up to 10.7B, including Qwen2.5-Omni-7B.

Getting started on a single GPU

All the models run through transformers: the latest version is enough for the text model, the VL model needs 5.21.0 or newer, and Audio and Omni need 5.20.2 or later, with the trust_remote_code flag. For local agents there is a separate Agent Harnesses guide, and instructions for llama.cpp and other engines are in the model cards.

The practical case is economic. A pipeline of ASR, VLM, LLM and TTS can be replaced with a single Omni model, which means fewer services, less latency at the seams and zero tokens sent to the cloud. A team with a narrow domain (medical terminology, warehouse jargon, internal interfaces) needs only one GPU to fine-tune a 2B model and deploy it on a laptop, a phone or a single-board computer. The best starting point is LFM3-1.7B for tool calling and structured outputs, adding the multimodal versions once a camera or microphone enters the input.