lfm3-1.7b-home-assistant

A fine-tune of LiquidAI/LFM3-VL-2B (LFM3 1.7B language model) on Home Assistant tasks: controlling devices, answering questions about the state of your home, managing to-do lists and saying no cleanly when it can't do something. A single 1.05 GB Q4_0 GGUF that runs on a Raspberry Pi 5.

Current version: v0.1 (lfm3-1.7b-home-assistant-v0.1.gguf). New versions are added to this repository as new files; the version is in the file name and in the GGUF metadata (general.version).

What it does

  • Acts on commands with the right tool and arguments: on/off, brightness and colour, covers and valves to a position, fans, media players, vacuums and locks, using the exact device, area and domain names your home exposes.
  • Answers questions about the current state of the home by reading the live context once, then giving one short sentence with the value and unit. It never changes a device while answering a question.
  • Handles to-do lists: add, read, count and complete items, picking the list that obviously fits and asking only when two lists fit equally.
  • Asks instead of guessing where a guess would be wrong or unsafe, for example an unnamed "lock the door" when there are several locks.
  • Refuses cleanly when a device or ability isn't exposed, and says what it can do instead.
  • Stays short: plain sentences without markdown, suited to voice. On ordinary questions it answers without calling tools.

How it scores against the competition

State as of 30 October 2026. Home LLM Leaderboard datasets, average weighted by number of samples as on the leaderboard. Other rows are as published; they were run on older Home Assistant versions with each integration's default settings, while v0.1 was run on Home Assistant 2026.10.3, so read the placement as indicative rather than as an official entry.

# model assist assist-mini questions average
1 muse-glimmer 90.2 % 98.0 % 97.6 % 94.3 %
2 gemini-3.1-flash-lite 87.0 % 98.0 % 100.0 % 93.8 %
3 gpt-5.4-mini 87.8 % 96.9 % 97.6 % 93.1 %
4 gemini-3.5-flash-lite 88.0 % 97.4 % 95.9 % 92.7 %
5 lfm3-1.7b-home-assistant v0.1 90.5 % 96.4 % 93.0 % 92.5 %
6 gemini-2.5-flash 88.7 % 96.9 % 94.6 % 92.4 %
7 gemma4-26b-a4b 86.3 % 98.0 % 97.0 % 92.4 %
8 gemma4-12b 83.5 % 97.4 % 100.0 % 92.1 %
9 minimax-m2.5 88.5 % 98.0 % 93.2 % 92.0 %
10 qwen3-30b-a3b-instruct 83.9 % 97.4 % 97.0 % 91.2 %
20 qwen3.5-4b 81.1 % 88.3 % 93.8 % 87.0 %
39 LiquidAI LFM3-VL-2B QAD Q4_0 (base model) 55.9 % 80.1 % 76.2 % 67.8 %

Ranks are among the 44 published entries plus this model and its base model. Its assist score (device control across many homes) is the second highest on the board.

Limitations

  • English only.
  • Some general knowledge is traded for the specialization (MMLU-Pro 23.3 % vs 26.4 % for the base model).
  • Weakest on multi-step reasoning over tool results, for example counting calendar events across a week.
  • Like any small model it can misread or mis-target. For locks, alarms and other security devices, keep Home Assistant's own confirmations and exposure settings in place.

License and modification notice

This model is a modified version of LiquidAI/LFM3-VL-2B and is distributed under the LFM Open License v1.0 (see LICENSE), including its commercial-use limitation (commercial use is licensed only for entities below US$10 million annual revenue). Modified: the language-model weights of LFM3-VL-2B were fine-tuned for Home Assistant tasks and re-exported as Q4_0 GGUF; the GGUF metadata name and version were changed. "LFM" and "Liquid" are names of Liquid AI, used here only to describe the origin of the model.

Downloads last month
302
GGUF
Model size
2B params
Architecture
lfm3
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Slomin/lfm3-1.7b-home-assistant

Finetuned
(23)
this model