LFM2.5-350M-home-assistant-sft

LFM2.5-350M-home-assistant-sft

Turns plain home commands such as turn off the TV into structured tool calls, checking device states first. A 350M model for smart home builders and tinkerers.

Made by OrdenWills on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a small smart home model, fine-tuned by OrdenWills from LiquidAI's LFM2.5-350M. It takes a spoken-style command plus the current state of your devices and replies with the right tool call. If nothing needs doing, it explains that in plain text. The card calls it a preview release.

Who it suits: It suits smart home tinkerers and developers who want a tiny model for device control. It covers TVs, speakers, fans, doors, lights, thermostats and whole-home scenes via tool calls.

Industry | Who uses it | How they use it | Example
Smart home | Home Assistant users | Turn text commands into device actions | Turn off what is on, and only the lit rooms get a call
Software | App developers | Use a small model for structured tool calls on modest hardware | Emit tool calls from a command and a device list
Hobby | Local AI fans | Run a model of 353 million parameters at home | Test undo and multi-device commands

People who want a general chat model can skip it, since it was trained for home control.

Where to use it:

• A smart home setup, such as the author's modified Home Assistant project on GitHub
• Any app that loads GGUF files, since the repository ships GGUF and Safetensors
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Device control | Home owners | A tool call for the right light, door, fan, TV or speaker
Undo and follow-ups | Developers | A reversing call, worked out from a log of recent actions
Impossible requests | Builders | A clear unclear-intent reply when a device is not connected

How to set it up: You need a few things first.

• Python and pip, for the Hugging Face hub tool
• A runtime for GGUF files, or the author's Home Assistant project
• About 177 MB for a 4-bit copy, by estimate, or about 707 MB in F16

1. Install the hub tool with pip.
2. Download the repository files.
3. Pick one of the GGUF files, which the card says are the most stable.
4. Load it in your runtime or in the author's Home Assistant setup.

pip install -U huggingface_hub
hf download OrdenWills/LFM2.5-350M-home-assistant-sft

Title: Which GGUF file to choose
File type | Use
F16 | The full 16-bit file
Q4_K_M | The card's pick for most people, trading a little detail for speed and size
Q8_0 | The best quality of the three

How to use it: You send the model a plain command together with the current states of your devices, and it answers with a tool call or a short text reply. The card shows the tool names below in its examples.

Tool | What it does
toggle_lights | Switches the lights in a room on or off
control_speaker | Plays media on a speaker in a room
intent_unclear | Reports that a request cannot be done, with a reason such as an unsupported device

Example: You say "Turn off what's on" and pass in the device states. The model reads them and emits one tool call for each room that has a light on. If every light is already off, it replies in plain text and makes no tool call.

For newcomers and experts: If you are new, think of it as a translator from everyday sentences to device commands. Download a GGUF file, then try a simple request like turning off one light and see the tool call it writes.

Underneath, the model reasons in a short think trace and ends it with a fixed phrase, either action required or action not required, before it opens a JSON block. For compound commands it counts the calls first and emits them all in parallel. The card reports 98.5 percent overall accuracy over 59 edge-case categories in its V14 test. It scores 52.0 percent on thermostat minimum and maximum cases and 84.0 percent on contextual log follow-ups.

Command and device states go in
Model reasons in a short think trace
Action required or not is decided
Tool call or plain text comes out

Good to know:

• License: Apache 2.0.
• Cost to run: no API key or paid service named. The card links a Home Assistant setup you would run yourself.
• Data it reaches: the device states and recent actions you pass in. Wired to a smart home, its tool calls can operate real devices.
• Not for: general chat, or relying on the Safetensors build, since the card calls it a preview and advises the GGUF files for stability.

Title: Accuracy in sample categories, V14
thermostat_min_max: 52.0%
contextual_log_followups: 84.0%
current_track_queries: 100.0%
mixed_device_undo: 100.0%
speaker_volume_controls: 100.0%

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 353M
• Weights: about 707 MB in F16; about 177 MB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 2,229
• Likes: 8
• License: apache-2.0
• Languages: en
• Based on: LiquidAI/LFM2.5-350M
• Tags: safetensors, gguf, lfm2, smart-home, home-automation, tool-calling, function-calling, unsloth, lora, text-generation, conversational

Links

• Model: https://huggingface.co/OrdenWills/LFM2.5-350M-home-assistant-sft
• Files: https://huggingface.co/OrdenWills/LFM2.5-350M-home-assistant-sft/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OrdenWills/LFM2.5-350M-home-assistant-sft

Then follow the usage in its model card. Nothing in this kit runs by itself.