Models Published: Oct 27, 2026

Liquid AI releases LFM3: a 1.7B agent model with 256K context that fits in about 1 GB

What happened

Liquid AI released LFM3 on October 27 as open weights on Hugging Face: a 1.7B text model with a 256K-token context window, plus vision, audio and omni models of about 2B parameters and 2B decision models (d1). LFM3 uses a new tri-hybrid architecture that mixes short convolutions, sliding-window attention and global attention, and was pre-trained on 22T tokens. Liquid says the text model needs about 1 GB at 4-bit, and that its full 128K-token KV cache takes 1.6 GB against 4.3 GB for Qwen3.5-4B. Ollama v0.40.12 added the models the same day, on Apple hardware first.

Why it matters locally: A tool-calling model with long context that fits in about 1 GB can run agents on ordinary laptops and phones, though the LFM Open License limits free commercial use to companies under US$10 million in annual revenue.

Three ways to read it

For the market

Liquid compares LFM3-1.7B with Qwen3.5-4B, Qwen3.5-9B and Gemma 4 E4B and reports higher tool-calling scores (71.39 on BFCL V4 against 62.83 for Qwen3.5-9B), while admitting larger models still win on knowledge and agentic coding. Its licence is the catch: unlike Apache-licensed rivals, larger companies need a separate commercial agreement, which may slow enterprise adoption.

In plain words

Liquid AI released a very small AI model that can use tools, read long documents and run on a phone or laptop without the internet. Think of it as a quick, well-organised assistant rather than an expert: good at following steps, less good at deep knowledge.

For practitioners

Liquid publishes GGUF and MLX builds and lists llama.cpp, Ollama, LM Studio, vLLM and SGLang, but llama.cpp needs an LFM3-compatible build and Ollama v0.40.12 supports lfm3 and d1 2B only on Apple hardware for now, with Linux and Windows to follow. Test it on your own tool schemas and long-context retrieval before relying on it: Liquid's own RULER 128K score is 65.56 against 87.11 for Qwen3.5-9B.