LFM3-1.7B

LFM3-1.7B is Liquid AI's post-trained, text-only member of its LFM3 open-weight family. It is designed for specialized local applications and agent harnesses where a small model, long context, tool use, and predictable deployment formats matter more than frontier-scale reasoning. Liquid AI released the family on October 27, 2026; the model card lists 1.73B parameters, a 256K-token input-and-output context window, and the LFM Open License v1.0.

What it is for

The checkpoint is trained for agentic harnesses including Hermes Agent, Pi, and OpenClaw. Liquid AI describes it as suitable for tool calling, structured outputs, and domain-specific fine-tuning. Its tri-hybrid architecture combines gated short convolutions, sliding-window attention, and global attention. The model card lists sixteen languages, including English, Chinese, and Japanese.

For deployment, official releases cover BF16, FP8, NVFP4, GGUF, MLX, and ONNX variants. That makes it practical to test the same model with Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, or ONNX Runtime according to the hardware available. Liquid AI says the 4-bit form needs about 1 GB of memory; treat that as a vendor estimate and measure your own prompt length, concurrency, and runtime overhead.

A practical workflow

1. Start with the official GGUF, MLX, or ONNX build that matches the target device.
2. Give the model one read-only tool and validate JSON or schema-constrained outputs on representative tasks.
3. Add retrieval for knowledge-heavy questions, as the model card recommends.
4. Fine-tune only after preserving an evaluation set that includes tool errors, ambiguous inputs, and multilingual examples.

Limits and trade-offs

LFM3-1.7B is not presented as a coding agent or a complex multi-step reasoning model. Liquid AI explicitly recommends against those uses because of its small size. A long context limit does not guarantee reliable recall, planning, or factual accuracy. Keep tool permissions narrow, require confirmation before any write or external action, and compare end-to-end task success against a larger model before production rollout.

Alternatives

• LFM 2.5-2.6B is Liquid AI's earlier small agent-oriented model.
• Qwen 3.8 27B is a larger open-weight option when local hardware permits it.
• Gemma 3 is Google's open model family for on-device work.

Conclusion

Choose LFM3-1.7B when a compact, locally deployable model needs structured output and constrained tool use. Its broad official packaging lowers deployment friction, but its intended scope is specialized automation, not a substitute for larger coding or reasoning models. Begin with the official model card and verify the model, license, and runtime path in your own environment.