Liquid AI splits LFM3 into six open 2-billion-parameter models built for the edge
On 27 October 2026, Liquid AI released LFM3, a family of six open-weight models of roughly 2 billion parameters — text, vision, audio, omni and decision — all branching from one shared backbone. The bet is explicit: run multimodal AI on a phone, a Raspberry Pi or a Jetson, with no cloud in the loop.
27 October 2026. Liquid AI releases LFM3, a new generation of open-weight Liquid Foundation Models. Six models of roughly 2 billion parameters — text, vision, audio, omni and decision — all start from the same backbone, LFM3-1.7B-Base. 27 October 2026. The weights land on Hugging Face, the Liquid AI playground and its API, under the LFM Open License v1.0. Why it matters: a frontier lab is pushing the “one model, many modalities” logic to its end — and explicitly targeting on-device execution, where the big models do not go.
One backbone, six specialisations
The family fans out from a single trunk. LFM3-1.7B takes text and returns text, with 1.73 billion parameters, a 256,000-token context window, a 128,000-token vocabulary and support for 16 languages. Around it sit three multimodal variants: LFM3-VL-2B (text, image and video in, text out), LFM3-Audio-2B (text or audio in, text or speech out) and LFM3-Omni-2B, the company’s first omni model, which takes text, image, video and audio and answers in text and speech. Two decision models, d1-2B and d1-omni-2B, are specialised from the vision and omni branches: they return calibrated answers — a list of questions answered in a single forward pass with probabilities — rather than generated prose.
The launch partners say it all. The list mixes silicon and edge ecosystems — Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, Qualcomm via AI Hub and GenieX — with OpenRouter and Vercel. Mercedes-Benz and Shopify took part in pre-release testing. The message is unmistakable: this model is designed to be deployed close to the user, not inside a data centre.
A tri-hybrid architecture tuned for embedded silicon
The LFM3-1.7B-Base backbone has 24 layers: 13 gated short convolutions (kernel size 7, up from 3 in LFM2), 5 sliding-window attention layers (a 256-token window) and 6 global attention layers. Three choices define the backbone: head-wise gated attention, where a sigmoid gate scales each head’s output and can suppress the activation outliers that hurt quantisation; partial RoPE, where global attention layers rotate only the highest-frequency half of each head while sliding-window layers keep full RoPE; and a sandwich norm that normalises each block’s input and output.
The design comes from hardware-in-the-loop architecture search extended to 14 device-and-backend combinations: phones (Samsung Galaxy S26 Ultra, OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max, AMD Ryzen AI), Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU. This is what sets Liquid AI apart from most labs: the architecture is optimised on the target silicon, not just on a training cluster.
Training follows the same logic. The backbone was pre-trained on 22 trillion tokens, the final 3.5 trillion of which were a mid-training stage on curated data with a larger share of agentic traces and tool calls that stretched the context to 256,000 tokens. Post-training runs in four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning, multi-domain on-policy distillation from six specialist teachers, then agentic RL inside real agent harnesses.
What the numbers are worth — and what to take from them
The published performance is, as ever, vendor-supplied. It deserves the same reading as a spec sheet: an order of magnitude, not a verdict.
LFM3-1.7B reportedly beats its predecessor LFM2.5-2.6B with 35% fewer parameters, and matches LFM2-1.2B per parameter on short prompts while running about 25% faster at 8,000 tokens of context. At 4-bit precision the text model fits in about 1 GB. On WebArena-Verified it jumps from 3.02 (LFM2.5-2.6B) to 38.42 — the largest gains on agentic tasks, even though the model card states this size is not recommended for agentic coding or complex multi-step reasoning.
The vision branch is the most striking. LFM3-VL-2B is the first Liquid vision model to accept video, and Liquid AI reports an average of 71.4 across 62 vision benchmarks, up from 57.6 for LFM2.5-VL-3B, with 1.9 billion parameters against 3.1 — putting it on par with Gemma 4 12B, a model six times larger. A new in-house 164-million-parameter vision encoder uses 3× less compute than SigLIP2 and reportedly matches Qwen3.5-2B’s vision encoder with 2.2× fewer FLOPs.
Audio and omni follow the same arc. LFM3-Audio-2B supports 18 languages, posts a UTMOSv2 of 3.83 against 3.22, a time-to-first-token as low as 50 ms, and a LiquidCodec tokenizer that generates speech at roughly 87× real time on a single CPU core of an M4 Max. LFM3-Omni-2B (2.2 billion) reportedly beats models up to 10.7 billion parameters — including Qwen2.5-Omni-7B — on Daily-Omni, WorldSense and AV-SpeakerBench. The decision model d1-2B replaces d1-3B with a third fewer parameters and improves from 64.0 to 75.5 on ImajevBench.
# Serve LFM3-1.7B locally with llama.cpp (4-bit quantisation ≈ 1 GB)
llama-server -m LFM3-1.7B-Q4_K_M.gguf --ctx-size 32768 --port 8080Single-GPU fine-tuning, and robotics in the wings
Liquid AI stresses one practical point: every model fine-tunes on a single GPU, with notebooks covering SFT, DPO, GRPO and continued pre-training. The in-house example is telling: LFM3-VL-2B fine-tuned on about 3,000 public wind-turbine blade photos, over a few hours on an A100, lifts defect classification to 67% and detection F1 from 0.003 to 0.11. Modest in absolute terms, but exactly the kind of industrial task that large models cannot justify economically.
Robotics is the natural extension. Three models — orchestration, navigation, manipulation — are derived from LFM3-VL-2B. The manipulation vision-language-action model, trained across 15 embodiments, hits 96.2% average success on the four LIBERO suites and 45.7% on RoboCasa-GR1, generating action chunks in under 50 ms on a Jetson AGX Orin in TensorRT FP16. On throughput, LFM3-1.7B is said to produce 3.6 billion tokens per day on a single NVIDIA B300 at 1,024 concurrent requests. Checkpoints ship in BF16, FP8, NVFP4, GGUF, MLX and ONNX, with native support for llama.cpp, MLX and vLLM.
Verdict
If you want multimodal AI on-device — a local assistant, robotic perception, document triage at the edge — test LFM3 now: it is one of the few backbones covering text, vision, audio and decision in a 2-billion-parameter format that fits in 1 GB at 4-bit and fine-tunes on a single GPU. If you are after complex reasoning or agentic coding, look elsewhere: the model card itself advises against this size for those tasks, and a model this small has a hard ceiling that vendor benchmarks do not circumvent. Either way, measure on your own data before believing the charts: Liquid AI’s numbers are orders of magnitude, not verdicts — the only test that matters is your workload on your hardware. The durable signal of LFM3 is not one more benchmark; it is that the “big model in the cloud, small model at the edge” boundary is dissolving from below.