Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models

LFM3 is a family of six open-weight models in the 2-billion-parameter class, all branched from one shared base. They cover text, vision with video input, speech, a combined omni model, and two decision models. Liquid says each can be fine-tuned on a single GPU and the text model fits in about 1 GB of memory at 4-bit, which puts it on phones and laptops.

The headline claim is agent performance. On Liquid's own evaluations, LFM3-1.7B beats the much larger Qwen3.5-9B on function calling and web-navigation benchmarks. If that holds up, capable agents no longer need a data center round trip for every step.

We like where this is going. Small models that run on the device are good for privacy, good for cost, and good for anyone building in places with weak connectivity. A model that can call tools and navigate interfaces while living on a laptop changes who gets to build agents and what data has to leave the machine. Liquid also tuned the architecture against 14 device and backend combinations, including a Raspberry Pi 5 and a Qualcomm NPU, which is the kind of unglamorous engineering that actually gets models shipped.

Now the caveats, and they are real. The comparison numbers come from Liquid's own evaluation runs, and in at least one case its score for a rival differs from what that rival reported. Independent testing will matter. The model is also candid about weak spots: knowledge-heavy questions and agentic coding still favor bigger models, and long-context retrieval at 128K tokens trails Qwen3.5-9B by a wide margin. A small agent that is fluent at following instructions but thin on knowledge can confidently do the wrong thing, which is the risk I'm watching as these get wired into real workflows.

The license deserves attention too. These are open weights, but free commercial use stops at $10 million in annual revenue. That is generous for startups and hobbyists, and a meaningful constraint for larger companies comparing against Apache-licensed alternatives like Qwen, Gemma, and MiniCPM.

Our read: LFM3 is a serious entry in the on-device race, especially for tool-using agents where it posts its strongest results. Teams should treat it as a specialist, pair it with search or retrieval where knowledge matters, and verify the benchmark claims on their own tasks before trusting it with anything consequential.

Liquid AI's models use a hybrid design rather than a standard transformer. The LFM3 backbone mixes short convolutions with sliding-window and global attention layers across 24 layers. The base model was trained on 22 trillion tokens, with the final stretch focused on agent-style data and extending context to 256K tokens.

Weights are available on Hugging Face, with formats including GGUF, MLX, and ONNX, and support across runtimes such as llama.cpp, vLLM, SGLang, and Ollama.