LFM2.5-24B-A2B: Can Liquid AI's Model Run a Local Home Assistant Agent?
- News
- Rocks on Galaxy
- Tech
- 19 Oct, 2026
- 0
Can a model with 24 billion parameters run your smart home from a box on your desk, without a single cloud call? Liquid AI thinks so. On October 13 it released LFM2.5-24B-A2B, an open-weight reasoning model that activates only about 2 billion parameters per token and was trained specifically to call tools inside agent frameworks. That combination, a large model that behaves like a small one at inference time, is exactly what a local Home Assistant agent needs.
We went through the model card, the published files and Liquid's own benchmarks to answer the practical questions: what changed, where it is genuinely strong, how much memory it really takes, and how to wire it into Home Assistant today.
LFM2.5-24B-A2B at a glance, versus its predecessor
| Specification | LFM2-24B-A2B (previous) | LFM2.5-24B-A2B (new) |
|---|---|---|
| Total / active parameters | 24B / about 2B | 24.0B / 2.45B |
| Architecture | 40 layers (30 gated short-convolution blocks + 10 attention), 64 experts per MoE block, top-4 routing | |
| Context window | 32,768 tokens | 131,072 tokens |
| Vocabulary | 65K tokens | 128K tokens |
| Pre-training data | About 17T tokens | About 33T tokens |
| Answer style | Not reasoning-only | Always reasons first (<think> block) |
| Languages listed | — | 16, including English, French, Spanish, Italian, German and Polish |
The architecture is unchanged; the gains come from data and training. Pre-training roughly doubled, two mid-training phases added reasoning and long-context data, and post-training ran in five stages ending with reinforcement learning inside real agent harnesses: OpenClaw, Hermes Agent, Pi and OpenCode. In other words, the model has practised using the same tool loops you would run at home.
Where it wins, and where it does not
Liquid compares the model with three mixture-of-experts rivals that have more active parameters. These are the vendor's figures, run with vLLM; we have not reproduced them.
| Benchmark (Liquid's figures) | LFM2.5-24B-A2B | Qwen3.6-35B-A3B | Gemma-4-26B-A4B-it | Nemotron-3.5-Lightning 30B-A3B |
|---|---|---|---|---|
| AutomationBench (agentic) | 27.69 | 8.72 | 5.46 | 2.61 |
| LiveClaw (agentic) | 48.09 | 38.19 | 41.25 | 37.72 |
| ToolSandbox (tool use) | 80.71 | 75.02 | 72.09 | 75.13 |
| BFCLv4 (function calling) | 65.01 | 65.00 | 57.09 | 53.78 |
| IFEval (instruction following) | 96.75 | 93.06 | 92.00 | 94.19 |
| GPQA Diamond (knowledge) | 74.42 | 83.69 | 74.90 | 72.42 |
| SWE-Bench Verified (coding) | 51.00 | 66.20 | 25.80 | 53.40 |
| WebArena Verified (browsing) | 26.75 | 37.62 | 25.62 | 28.69 |
- Strong: tool calling and following instructions, the two skills that matter when an agent has to turn "dim the living room and close the blinds" into the right calls. Liquid says it leads or ties on five of eight agentic benchmarks and on all three instruction-following tests.
- Weaker: raw knowledge, coding and web browsing, where the larger Qwen3.6 keeps a clear lead. Liquid itself advises pairing the model with retrieval or web search for knowledge-heavy work.
- An honest footnote: during training the model showed reward-hacking behaviour, looking up answers in a repository's git history or on GitHub, PyPI and Hugging Face, and even working out it was in a sandbox. Liquid says it tightened its sandboxes and that all published scores come from sealed environments. It is a useful reminder to give any agent only the permissions it needs.
Will it run on your machine?
Total parameters decide how much memory you need; active parameters decide how fast it answers. Here are the official files on Hugging Face and in the Ollama library:
| Format | File size | Use it with |
|---|---|---|
| GGUF QAD Q4_0 (quantization-aware distilled) | 13.58 GB | llama.cpp, LM Studio |
| GGUF Q4_K_M | 14.53 GB | llama.cpp, LM Studio |
Ollama lfm2.5:24b | 15 GB | Ollama, Home Assistant's Ollama integration |
| GGUF Q5_K_M / Q6_K | 17.03 / 19.69 GB | llama.cpp |
| GGUF Q8_0 | 25.49 GB | llama.cpp |
| MLX 4-bit | About 13.6 GB | Apple Silicon Macs |
| BF16 (full precision) | 47.96 GB | vLLM, SGLang, Transformers |
On speed, Liquid reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, which it says is up to 1.7 times faster than comparable MoEs while using the least memory.
Our rule of thumb, not a measurement: the 4-bit file alone is close to 15 GB, and the operating system plus the context cache need room on top. A 16 GB computer is too tight. Aim for 24 GB of unified memory or RAM as a floor and 32 GB to be comfortable. A Raspberry Pi is out of the question for the model itself, although it is still fine for Home Assistant; our Raspberry Pi 5 kit review covers that side. A gaming laptop like the HP Omen Max 16 we reviewed, with 32 GB of RAM and an RTX 5070 Ti, has enough total memory; llama.cpp can split a model between GPU and system RAM, at a speed we have not measured yet.
Key takeaways
- Open weights, 24B total but about 2B active: big-model skills, small-model speed.
- Built for agents: best scores on tool use and instruction following, behind Qwen3.6 on knowledge and code.
- About 14 to 15 GB at 4-bit: plan 24 GB of memory minimum, 32 GB for comfort.
- Works with Home Assistant today through Ollama (
lfm2.5:24b).- Free for companies under $10 million in annual revenue.
How-to: plug it into Home Assistant in five steps
- Install Ollama on the machine with the memory, a desktop, Mac or mini PC, not on the Home Assistant box if that is a Pi.
- Pull the model: run
ollama run lfm2.5:24band test a few prompts locally. - Expose Ollama to your network so Home Assistant can reach it, as the Home Assistant Ollama documentation requires.
- Add the Ollama integration in Settings › Devices & services, enter the server URL and the model name. Home Assistant sets an 8K context window by default, larger than Ollama's 2K default; raise it if you expose many entities.
- Expose only a few entities. Home Assistant recommends fewer than 25 when experimenting with local models, and only models that support tools can control the house. A tip from the same page: create two Ollama entries with the same model, one for conversation and one with control enabled.
If you prefer an agent framework, the route is Home Assistant's Model Context Protocol Server: enable it under Settings › System › AI and point an MCP-capable agent at it. Ollama's LFM2.5 page lists one-line launchers for Hermes Agent and OpenClaw, and both harnesses were part of this model's training. We covered the one-click MCP server and the other autumn changes in our Home Assistant 2026 guide.
Home Assistant compatibility: our notes
- It always thinks. The model card says LFM2.5-24B-A2B reasons before every answer. That helps multi-step requests but adds tokens before the reply. At Liquid's 184 tokens per second, every 200 reasoning tokens add roughly a second; for simple "turn off the light" commands, Home Assistant's built-in sentence matching remains faster.
- Use the recommended settings: temperature 0.1, top-k 50 and a repetition penalty of 1.05. The GGUF files carry them as metadata, so llama.cpp applies them automatically.
- Pair it with a source of facts, a search tool or your own documents, rather than asking it general-knowledge questions cold.
- Local hubs are going the same way. Several 2026 hubs promise on-device AI; see our guide to Matter hubs with local AI. If you are still building the sensor network this agent will control, our analysis of the Zigbee vs Matter over Thread lab study will help you pick a protocol.
Who is it for?
- Yes: Home Assistant tinkerers who want a private agent that can plan and call tools, households that speak French, Spanish, Italian, German or Polish as well as English, and anyone who already owns a 32 GB machine.
- Not yet: owners of 16 GB laptops, people who mainly want a chatbot for general knowledge, and anyone expecting instant replies to one-word voice commands.
- Thinking of buying hardware for it? Wait for the November sales and compare memory first; our Black Friday 2026 planning guide explains how to spot a real discount.
License and cost
The weights are published under the LFM Open License v1.0. Personal and research use is free, and commercial use is allowed as long as your organisation stays below $10 million in annual revenue; above that threshold, commercial use is not covered. Qualifying non-profits are exempt for non-commercial and research purposes. If you would rather not host it, Liquid's API charges $0.10 per million input tokens ($0.05 cached) and $1.00 per million output tokens, with $2 of starting credit.
FAQ
What does "A2B" mean?
It stands for about 2 billion active parameters. The model holds 24 billion in total, but its mixture-of-experts design only uses four of 64 experts per block for each token, which keeps it fast.
Can it run on a Raspberry Pi with Home Assistant?
No. Even the smallest official file is 13.58 GB. Keep Home Assistant on the Pi and run the model on another machine reached through the Ollama integration.
Does it speak French, Spanish and Italian?
Yes. Liquid lists 16 supported languages, including those three. The vocabulary was also doubled to 128K tokens, mainly to handle non-Latin scripts such as Arabic, Hindi or Thai better.
Is it better than Qwen3.6-35B-A3B?
For tool use and instruction following, Liquid's benchmarks say yes; for knowledge, coding and web browsing, no. For a home agent, the first group matters more.