LFM3-1.7B is LiquidAI’s small text model for local agents and specialized applications. Its useful distinction is a deployable 1.73B-parameter checkpoint with tool-oriented training and several local runtime routes. It is not a miniature replacement for a frontier coding agent: the vendor explicitly advises against agentic coding and complex multi-step reasoning with this model.

This guide covers the released text checkpoint, rather than every model in the LFM3 family. We have inspected documentation and model artifacts, not run an independent speed, battery or accuracy benchmark. Check the official model card alongside the runtime recipe you intend to use.

What the checkpoint actually contains

Property	Documented specification
Checkpoint	LiquidAI/LFM3-1.7B
Parameters	1.73B
Input and output	Text
Context	256K tokens, shared by input and output
Architecture	24-layer tri-hybrid
Layers	13 gated short convolution, 5 sliding-window attention, 6 global attention
License	LFM Open License v1.0

The family includes other modalities, but this checkpoint does not become a vision or audio model because its siblings support them. Choose the exact repository and record its revision. A family name is not a sufficient deployment configuration.

LiquidAI describes training for Hermes Agent, Pi and OpenClaw-style harnesses. A compatible tool-call format is an integration feature, not permission for the model to execute a shell. The harness still supplies tools and determines which actions are allowed. Our agent security guide explains that boundary.

Local runtime choices

LiquidAI lists Transformers, vLLM, SGLang and LFM3-compatible llama.cpp builds. GGUF artifacts can be used through the documented llama.cpp, Ollama and LM Studio routes. Support for earlier LFM releases does not establish support for this architecture in an old binary. Use the deployment documentation and checkpoint-linked instructions.

A practical setup sequence is to select a runtime, verify its LFM3 support, download the corresponding checkpoint format, pin the model revision and start with a short prompt. Confirm that the model uses the current chat template before enabling tools or long context. Template mismatches can look like weak reasoning or malformed tool calls even when the underlying weights loaded correctly.

The official Transformers example enables trust_remote_code=True. That permits downloaded repository code to run locally. Inspect and pin that code before using it on a machine containing credentials. A model download and executable loader code deserve separate review.

For interface choices, compare our Ollama guide and LM Studio review. Neither guide guarantees that every installed version recognizes this checkpoint. The Hermes local guide covers the surrounding persistent-agent process.

Memory: a small model is not a fixed application budget

LiquidAI says the model occupies about 1 GB at four-bit precision. Treat that as the vendor’s model-footprint claim, not a measured peak memory requirement for your application. The runtime, prompt processing, context cache, tools and concurrent sessions add memory. A phone that can load the weights may still struggle with a long request or sustained generation.

The advertised 256K window is an upper capacity specification. It is not evidence that full-window use is fast, economical or useful on your laptop. Begin with the context your task requires, then increase it while measuring peak memory, first-response latency and successful completion.

LiquidAI’s published local comparisons use stated hardware, quantization and a 32,768-token context. Do not transplant those numbers to the maximum context or another device. Our GPU memory planning guide separates weights from the rest of the serving footprint.

Hosted API pricing is a different route

The OpenRouter model listing, checked November 7, lists liquid/lfm-3-1.7b at $0.08 input and $0.20 output per million tokens, with $0.02 cached input where supported. The listing advertises 262,144 context tokens and an 8,192-token maximum completion on the current top-provider record.

For example, 100,000 uncached input tokens and 10,000 output tokens total $0.01 at those rates. That is arithmetic, not a measured agent session. Tools, retries and other services can add cost. Do not describe these as LiquidAI direct API prices.

A separate liquid/lfm-3-1.7b:free route currently has zero listed token prices. Free-route quotas, availability and provider terms still apply. It is not an unlimited service commitment or evidence that local hosting has no hardware cost. The OpenRouter guide explains why routing and account access matter.

Commercial use needs a license check

The LFM Open License v1.0 contains commercial-use conditions and a $10 million annual-revenue threshold. Read the complete definitions and conditions for your organization, distribution and use. Do not label the weights Apache 2.0 or assume that downloadable means unrestricted commercial use.

Model weights, runtime code and any agent plugins can each have separate licenses. Record what you deploy instead of relying on a library’s package license to describe the model. Hosted API terms are another arrangement and do not automatically grant redistribution rights.

Useful tasks and limits

Good evaluation candidates are bounded extraction, formatting, routing and simple tool selection with a small, explicit schema. Validate the returned fields before using them. For factual questions, LiquidAI recommends retrieval or search; a small local model is not a continuously updated knowledge source.

Test a representative task set before handing it a recurring responsibility. Include ambiguous requests, missing information, invalid tool arguments and cases that require escalation. Count accepted outputs and corrections, not only tokens per second. A fast incorrect action can cost more than a slower model call.

My take

LFM3-1.7B is worth evaluating when local execution, a small weight footprint and narrow application control matter. I would start with one low-impact structured task and a restricted tool set. I would keep complex coding and open-ended autonomous work on a model demonstrated to handle those jobs, rather than treating the agent-oriented training label as a general capability guarantee.

FAQ

Is LFM3-1.7B a coding model?

It can participate in agent harnesses, but LiquidAI does not recommend it for agentic coding or complex multi-step reasoning. Test the actual task rather than relying on the harness name.

Does the 1 GB claim include a 256K session?

No complete application-memory guarantee follows from that weight-footprint claim. Context, runtime overhead and concurrency require separate measurement.

Can I use the free OpenRouter route in production?

Check its current quotas, service availability and applicable terms. Zero token pricing does not establish a production service commitment.

Is it unrestricted open source?

The model uses LFM Open License v1.0 with commercial conditions. Review that license separately from runtime code.