Liquid AI's LFM3 is a small open-weight family built for agents

Liquid AI released LFM3, a family of small open-weight models that share one backbone. The first wave centers on models around 2 billion parameters covering text, vision, audio, an omni model, and two decision models, all branched from the same LFM3-1.7B base. The pitch is specialized intelligence that runs locally: the text model fits in about 1 gigabyte at 4-bit precision, each model can be fine-tuned on a single GPU, and a full 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B.

The architecture is where the work shows. LFM3 uses a "tri-hybrid" design of 24 layers: 13 gated short-convolution layers, 5 sliding-window attention layers, and 6 global attention layers, found through hardware-in-the-loop search. On Liquid's benchmark table the 1.7B model posts an overall average of 49.37, ahead of Qwen3.5-9B at 47.24, and it is especially strong on agentic tasks like tool calling (BFCL V4 at 71.39) and instruction following (IFEval at 96.66). It is weaker where you would expect a small model to be, trailing larger models on coding (SWE-bench Verified 43.40 against 51.20) and knowledge (GPQA Diamond 63.37 against 80.40).

Liquid says the base model was pre-trained on 22 trillion tokens, with a final stage adding agentic traces and tool calls, then post-trained through supervised fine-tuning, mixed reinforcement learning, distillation from six specialist teachers, and agentic RL in real harnesses. One stress test ran more than 300,000 parallel agents over the Enron email archive. The details are in Liquid's announcement.

Why it matters

If you build on-device or high-volume agents, a 2B model that beats a 9B on tool calling while using a third of the memory changes what you can run cheaply. Test the agentic benchmarks against your own tasks, since the coding and knowledge scores still favor larger models.