There's a persistent assumption that Home Assistant is the powerful-but-dumb option: great at execution, no intelligence. That hasn't been true for a while, and repeating it does a disservice to anyone trying to make a decision.
Home Assistant can run a large language model as its conversation agent, entirely on your own hardware, today. That's worth understanding properly, including if you're evaluating us.
What works now
Home Assistant ships a native Ollama integration, so a model running on your own machine can act as the conversation agent with no custom glue code. You say "make the living room cosy" and the model maps that into concrete device calls: dim the lights, warm the colour temperature, lower the blinds. It isn't keyword matching. The model can weigh time of day, who's home and current sensor states against a natural-language goal.
The voice pipeline is local too. Assist chains speech-to-text, the conversation agent and text-to-speech over the Wyoming protocol, an open standard from the Rhasspy project. With Whisper for transcription and Piper for speech, the whole loop runs in your house.
Speed is usable now, which wasn't always true. Home Assistant reworked its text-to-speech to stream, so Piper and Home Assistant Cloud TTS start speaking as soon as the model produces its first few words. They measured that as roughly a tenfold improvement in spoken response time. On a GPU with 8 GB of VRAM, a 7B or 8B model typically responds in under two seconds.
Model choice has settled too. Community consensus in 2026 has landed on the Qwen3 family: roughly Qwen3 4B Instruct on 8 GB cards, Qwen3-VL 8B in the middle, Qwen3 30B on 24 GB cards.
That's a real local AI stack, free and open source. Credit where it's due.
What it costs you to get there
The capability is real. The path to it isn't free of effort.
You need hardware. A responsive local LLM wants a GPU, and a Raspberry Pi running Home Assistant won't do this well. That's a second machine, its power draw and its noise.
You need to assemble it: Ollama, a model choice, Whisper, Piper, the Assist pipeline, and the exposure settings that decide which entities the model is allowed to touch. Every piece is documented. There are still a lot of pieces.
The genuinely hard part is model behaviour, specifically picking one that calls Home Assistant's tools correctly instead of guessing at them. Smaller models hallucinate device names and invent capabilities. That's tuning work, and it's the step people underestimate.
Then there's maintenance. Models improve, integrations change, and what you assembled in spring wants revisiting by autumn.
None of this is a criticism. It's the normal cost of an open, composable system, and for a lot of people it's a pleasure rather than a chore.
Curious whether Nexop fits your home?
Book a live demo, run by one of the founders rather than a salesperson. Or join the waitlist and hear from us the day it ships.
The gap that remains
There's one place where we differ from all of that, and it's worth being precise about it.
Everything above is reactive. However sophisticated the model is, the interaction starts with you. You ask, it acts. A local LLM in Home Assistant is a much better interface to your automations, but it doesn't reduce how many automations you have to think of in the first place.
Nothing in that stack watches a week of your household and says you dim the lights and close the blinds around 20:30 most weeknights, want that automatic? You still supply the intent. The model is a superb translator of intent, not a source of it.
That isn't a flaw in Home Assistant, it's a different product goal. But if your problem is "I never got around to building the automations", a better command interface doesn't solve it. That's the proactive automation problem, and it's a different piece of engineering.
Which should you choose
Stay with Home Assistant if you enjoy assembling systems, you already have hardware that can host a model, or you want maximum control with no vendor in the loop. It's free, it's excellent, and its ceiling is far above ours. If you want the AI parts without the assembly, Home Assistant Cloud at $6.50 a month is a fair deal that also funds the project.
Consider Nexop if you want the local-first properties without building the stack, and specifically if you want the house to propose automations rather than wait to be asked. We support Zigbee, KNX, Matter and WiFi natively, with no dongle needed for Zigbee. But Home Assistant's integration library is vastly larger than ours will ever be, and we don't support Z-Wave or Thread at launch.
The fuller side-by-side is on our Home Assistant comparison, which also says plainly where Home Assistant is the better choice.
Nexop is pre-launch. If you want a local AI voice assistant this month, the Home Assistant stack above works and we'd rather point you at it than pretend otherwise. If the proactive part is what you've been waiting for, book a demo or join the launch list.