Liquid AI releases LFM3, six open-weight 2B-class models built for local agents

Liquid AI released LFM3, six open-weight ~2B models for text, vision, audio, omni and decisions, with the text model post-trained for agent harnesses.

Liquid AI released LFM3 on October 27, 2026. It is six open-weight models of about 2B parameters each, covering text, vision, audio, omni and decision tasks, all built on one shared backbone, according to Liquid's announcement. The text model, LFM3-1.7B, is post-trained to run inside agent harnesses such as Hermes Agent, Pi and OpenClaw. It has a 256K-token context and takes about 1 GB of memory at 4-bit, per its model card. If you want an always-on local agent on a laptop or phone, it is worth a test.

Weights are on Hugging Face now, with GGUF, MLX, ONNX, FP8 and NVFP4 variants. The model card lists llama.cpp, Ollama, LM Studio, vLLM, SGLang and Transformers as supported runtimes. It also documents XML-style tool calls and deferred "tool search" for large tool catalogs. Liquid says every model can be fine-tuned on a single GPU, and its Hugging Face post gives setup starting from pip install -U transformers.

The benchmark story is mixed, and all of it is Liquid's own. Liquid reports that LFM3-1.7B scores 38.42 on WebArena-Verified, against 28.94 for Qwen3.5-9B. It also leads on BFCL V4, 71.39 to 62.83. MarkTechPost notes that Qwen3.5-9B still wins on GPQA Diamond and SWE-bench Verified. Liquid's own model card says the model is not recommended for agentic coding or complex multi-step reasoning.

OrcaRouter points out that none of the numbers have been independently reproduced, and that Liquid publishes no context length for the vision, audio, omni or decision models.

Licensing matters for client work. The models ship under the LFM Open License v1.0, which MarkTechPost says allows free commercial use only below $10M in annual revenue.

Sources