Liquid AI LFM3: six open-weight 2B models on one 1.7B backbone for on-device agents

Liquid AI announced LFM3 on 27 Oct 2026: six open-weight models in the 2B class — text, vision-language, audio, omni, and two decision models — all branched from one 1.7B base pre-trained on 22T tokens. The 1.7B text model has a 256K context window and fits in about 1GB at 4-bit precision; weights ship on Hugging Face with GGUF, MLX, ONNX, BF16, FP8, and NVFP4 variants, plus day-0 runtime support from llama.cpp, MLX, vLLM, Ollama, and LM Studio.

Why it matters

What you could build with it

Does it hold up?

Usable on day one for local experiments: weights are downloadable with broad runtime support (llama.cpp, MLX, vLLM, Ollama), but benchmark claims are vendor-reported and the 1.7B text model is not recommended for agentic coding or complex reasoning.

Built with Liquid AI LFM3

• Liquid AI's LFM3: Six 2B Open Models, One Shared Brain — and a 1GB Phone-Sized Agentarticle · Independent walkthrough: the six-model lineup, 1GB footprint claim, LFM Open License terms, and the caveat that vendor benchmarks are unverified.
• Liquid AI Launches LFM3 Family of Six Open-Weight 2B-Class Modelsarticle · Unite.AI launch coverage: per-model input/output modalities, 256K context, 16 languages, and day-0 partners including OpenRouter, Qualcomm AI Hub, and Vercel.
• Liquid AI's LFM3 Shrinks Six Specialized Models to Run on Your Phonearticle · AlphaSignal analysis: a practical evaluation checklist — pick the modality, check the model card, benchmark the target device, then fine-tune against held-out data.

Learn more

official announcement →

First spotted on article: source.

More AI models