Liquid AI launches LFM3: six open-weight 2B-class models on one shared backbone

Liquid AI has a new model family, and its pitch is the opposite of the frontier's usual race to bigger. Announced on October 27, 2026, LFM3 packs six open-weight models into the 2B parameter range — covering text, vision, audio, omni, and decisions — all branched from one shared backbone. The idea is not one giant generalist but a small specialist you can fine-tune on a single GPU and run wherever your application lives: a phone, a laptop, or an edge device with no cloud round trip.

The base and post-trained models are open-weight and available now on Hugging Face, the Liquid AI Playground, and the Liquid AI API, under the LFM Open License v1.0. Liquid AI is positioning LFM3 as the starting point for what it calls specialized intelligence: models small enough to customize, fast enough to run locally, and capable enough to become reliable experts on one narrow task.

One backbone, six models#

Every LFM3 model branches from the same LFM3-1.7B-Base backbone, and all six sit around 2B parameters. Four cover input modalities: LFM3-1.7B takes text and returns text; LFM3-VL-2B accepts text, images, and video and returns text; LFM3-Audio-2B takes text or audio and responds with text, speech, or both; and LFM3-Omni-2B, Liquid AI's first omni model, accepts text, images, video, and audio and produces text and speech. The last two are decision models specialized from the vision and omni models: d1-2B reads text and pictures, and d1-Omni-2B also listens. Instead of generating an answer, they return calibrated decisions — each with a probability — for a whole list of questions in a single forward pass.

A tri-hybrid architecture built for the edge#

The backbone is a hybrid of three operators across 24 layers: 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. Three further changes define the design: head-wise gated attention, in which a sigmoid gate scales each attention head's output; partial RoPE, under which global attention layers rotate only the highest-frequency half of each head; and a sandwich norm that normalizes each block's input and output.

LFM3-1.7B-Base was pre-trained on 22T tokens, with the final 3.5T tokens forming a mid-training stage on curated data with a larger share of agentic traces and tool calls that also extended the context window to 256K tokens. The text model itself has 1.73B parameters, a 256K-token context length, a 128,000-token vocabulary, and support for 16 languages. At 4-bit precision, Liquid AI says, it takes about 1 GB of memory — small enough, the company claims, to fit on any phone or laptop. The architecture search behind it was expanded for LFM3 to 14 device and backend combinations, including phones, a Raspberry Pi 5, laptops, Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.

The company-reported numbers#

All of the performance figures below come from Liquid AI's own announcement and model cards; no independent benchmarks have been published yet. On agentic tasks, the company reports that LFM3-1.7B scores 38.42 on WebArena-Verified against 3.02 for its predecessor LFM2.5-2.6B, and 41.29 on AutomationBench against 0.94 — while using 35% fewer parameters than that predecessor. On a single NVIDIA B300 at 1,024 concurrent requests, Liquid AI says the text model can produce 3.6 billion tokens per day.

The vision model is the first LFM vision model to accept video input, and Liquid AI reports it raises the average across 62 vision benchmarks from 57.6 to 71.4 compared with LFM2.5-VL-3B, while using 1.9B parameters against 3.1B — putting it, the company says, on par with Gemma 4 12B, a model more than six times larger. Its new in-house 164M-parameter vision encoder uses 3x less compute than the SigLIP2 encoder in previous LFM vision models. The audio model now supports 18 languages, with speech quality of 3.8 on UTMOSv2 against 3.2 for LFM2.5-Audio-1.5B, time to first token as low as 50 ms, and a LiquidCodec tokenizer that generates speech at about 87x real time on a single CPU thread on an Apple M4 Max. The 2.2B-parameter omni model, Liquid AI says, beats models up to 10.7B parameters, including Qwen2.5-Omni-7B, on the Daily-Omni, WorldSense, and AV-SpeakerBench audio-visual understanding benchmarks.

The decision models get the most distinctive upgrade: typed answers. Beyond yes/no, choices, and scores, d1-2B and d1-Omni-2B can now point at objects (drawing a box around what you describe, or finding every instance with a count), read text, numbers, and dates out of a document with the exact span they came from, estimate a number with its likely range, and pick every option that applies, each with its own probability. Liquid AI says d1-2B replaces d1-3B with a third fewer parameters and scores 75.5 against 64.0 on ImajevBench.

Built to be specialized#

The through-line of the release is customization. Every LFM3 model can be fine-tuned on a single GPU, and Liquid AI published notebooks covering supervised fine-tuning, DPO, GRPO, and continued pre-training. In one company example, LFM3-VL-2B was fine-tuned on about 3,000 public wind-turbine blade photos for a few hours on one A100, raising defect classification accuracy to 67% and detection F1 from 0.003 to 0.11 on held-out photos.

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

Available now, with day-one partners#

LFM3 ships with native support for llama.cpp, MLX, and vLLM across Apple, AMD, NVIDIA, and Qualcomm hardware, with checkpoint variants including BF16, FP8, NVFP4, GGUF, MLX, and ONNX. Day-0 launch partners named by Liquid AI are 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.

The honest caveat is in Liquid AI's own model card: LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning because of its small size. That is the bargain LFM3 is offering, stated plainly — a 2B-class family that trades frontier generality for something the frontier cannot give you: a model you can own, specialize, and run anywhere, today.

Sources: Liquid AI's announcement, “Introducing LFM3: Every Modality for Specialized Intelligence” (October 27, 2026); the LFM3-1.7B model card on Hugging Face; Unite.AI's coverage of the LFM3 launch. The editorial illustrations are AI-generated.