Liquid AI shipped LFM3 on October 27 — six open-weight models, all around 2 billion parameters, covering text, vision, audio, omni, and two decision variants. The text model runs in about 1 GB at 4-bit precision. Every model is fine-tunable on a single GPU. All weights are on Hugging Face now. The size is the headline. The decision models are the actual story.
The Part Most Coverage Missed: Decision Models #
LFM3 ships two decision models — d1-2B and d1-Omni-2B — and they work differently from every other model in the family. They do not generate text. Instead, you define a set of questions, pass them an input (a support ticket, an invoice, an image, a voice note), and they return typed answers in a single forward pass: probabilities, choices with confidence scores, object pointers, extracted document values, number estimates with ranges.
No output parsing. No prompt engineering to coerce a JSON response. If you are building routing logic, document extraction pipelines, content moderation, or quality inspection workflows, this is a meaningfully different primitive from a chat model. d1-2B handles text and images. d1-Omni-2B adds video and audio input. Both are built on top of their respective VL and Omni backbone variants.
Liquid’s defect-classification benchmark is a useful reference point: fine-tuning LFM3-VL-2B on a few hours of training on one NVIDIA A100 took accuracy from under 10% to 67%. The fine-tune path to d1 is the same — pick a modality, fine-tune on your data, deploy locally.
The Six Models #
| Model | Params | Input | Output |
|---|---|---|---|
| LFM3-1.7B | 1.73B | Text | Text |
| LFM3-VL-2B | ~2B | Text, images, video | Text |
| LFM3-Audio-2B | ~2B | Text, audio | Text, speech |
| LFM3-Omni-2B | ~2B | Text, images, video, audio | Text, speech |
| d1-2B | ~2B | Text, images | Typed answers |
| d1-Omni-2B | ~2B | Text, images, video, audio | Typed answers |
All six share a single base — LFM3-1.7B-Base — built on a tri-hybrid architecture: 13 gated short-convolution layers, 5 sliding-window attention layers, and 6 global attention layers across 24 blocks. The convolutions handle local pattern matching cheaply. Sliding-window attention covers mid-range context. Global attention layers handle the long view at 256K tokens. Liquid says the architecture was found via automated hardware-in-the-loop search across 14 processor classes, which explains why it runs fast on both server GPUs and edge hardware.
Running LFM3 #
The text model in 4-bit quantization is approximately 1 GB — it fits on a phone. Liquid validated performance on Apple M5 Max, AMD Ryzen AI Max+, and iPhone 17 Pro. At 32K context, LFM3-1.7B prefills at least 2.5x faster and decodes at least 1.8x faster than all compared models on all three platforms, while using less memory.
Available in every major format: GGUF for llama.cpp and Ollama, MLX for Apple Silicon, BF16/FP8/NVFP4 for vLLM and SGLang, ONNX for ONNX Runtime, and standard Hugging Face Transformers.
The quickest path to running the text model:
ollama run liquidai/lfm3-1.7b
Or via the Hugging Face Python API:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM3-1.7B")
tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM3-1.7B")
Performance Against Larger Models #
LFM3-VL-2B outperforms all models of similar size by at least 7.4 points across 62 vision benchmarks and matches Gemma 4 12B — a model six times larger. LFM3-1.7B edges Gemma 4 12B on LiveCodeBench v6 (73.46 vs 72.0). Whether these benchmarks reflect your actual workload depends on what you are building, but for a 2B-class model the numbers are unusually competitive.
License and Where It Fits #
LFM3 ships under the LFM Open License v1.0. Companies under $10M in annual revenue get royalty-free commercial use. Larger organizations need to negotiate directly with Liquid AI. It is not Apache or MIT — worth checking before you ship to production.
The deployment story is strongest for teams that need on-device inference for privacy or latency, edge agents without cloud dependency, enterprise-controlled infrastructure, or structured inference pipelines where decision models replace brittle prompt engineering. Liquid provides fine-tuning notebooks for every modality. The shared backbone means specializing one variant does not require starting another from scratch. If you have been running LFM2.5-24B-A2B for agent workloads, the LFM3 text model is worth benchmarking as a drop-in replacement on constrained hardware.
Weights are live on Hugging Face. The full technical breakdown is on the Liquid AI blog, and the Hugging Face technical post covers fine-tuning specifics for each modality.