Open WebUI + Ollama Always-On Home Agent Box (Mac mini and Strix Halo)

By LLM Hardware team · Methodology ·

Key facts

Stack
Open WebUI in front of Ollama. No model bundled in the chat path.
24 GB mini
7B comfortable, 27B Q4 tight.
36 GB Studio / 128 GB EVO-X2
36 GB Studio for 27–32B. 128 GB EVO-X2 for large MoE.
Checked
2026-10-13

Checked 2026-10-13. Author: LLM Hardware team. Amazon prices are not stored or shown on this site.

Citeable facts

Key fact: Open WebUI is a self-hosted interface for Ollama and OpenAI-compatible APIs, including a fully offline install.

How we got it: docs.openwebui.com, read 2 October 2026. Docker, pip, uv, and a desktop app are all documented. The :main image bundles embeddings; :slim does not.

Checked:

Key fact: Mac mini M5 Pro 24 GB maps to a 7B comfortably and a 27B Q4 tightly (18.2 GB). Studio M5 Max 36 GB maps to 27–32B with headroom. EVO-X2 128 GB is the large-MoE tier.

How we got it: Q4 formula params × 0.5 × 1.2 + 2. ASINs B0HGGHNQY6, B0HGKSQMX6, B0F53MLYQ6. Prices checked 1 October 2026 and will move.

Checked:

Key fact: Open WebUI Computer is a separate, faster-moving agent harness from the same project. It is not required to run the chat UI.

How we got it: Same docs page. Computer uses approvals and plan mode on the machine you install it on. The stable multi-user product described there is Open WebUI itself.

Checked:

Key fact: Three simultaneous 8K chats on one always-on Ollama server need about 10.0 GB with Llama 3 8B, 22.7 GB with Qwen3-30B-A3B and 27.6 GB with Qwen3-32B, all at Q4.

How we got it: Formula result + 3 x f16 KV cache for 8,192 tokens; needs OLLAMA_NUM_PARALLEL=3 (Ollama FAQ). Mac usable memory after the catalog's 4 GB macOS reserve.

Checked:

Key fact: Apple lists the Mac mini M6 with 16GB and a 256GB SSD at 4 W idle and 70 W maximum. At idle around the clock that is 2.9 kWh in a 30-day month.

How we got it: Apple support article 103253 (Mac mini power consumption and thermal output), read 2026-10-30. 4 W x 720 hours / 1,000.

Checked:

Key fact: The Open WebUI README command for Ollama on the same computer reaches it through host.docker.internal and sets --restart always, which Docker documents as restarting the container when the Docker daemon restarts.

How we got it: github.com/open-webui/open-webui README; docs.docker.com, Start containers automatically. Read 2026-10-30.

Checked:

Key fact: In Open WebUI the first account is the administrator, and later sign-ups stay pending, with no access to models, until an admin approves them.

How we got it: docs.openwebui.com quick start and "Share Open WebUI with Your Team", read 2026-10-30.

Checked:

Methodology explains the VRAM formula, the bytes per weight at each quantization, and the full source list.

Open WebUI plus Ollama is the stack to install if you want a ChatGPT-shaped window on models you downloaded. Open WebUI’s docs call it a self-hosted platform that can run offline, with Ollama and any OpenAI-compatible API behind it. Ollama is the model runtime: pull a GGUF-style tag, serve it on localhost. Neither product is a cloud agent. Meta Muse, dots, and Grok Bot are the cloud products this box is an alternative to. The comparison is on the hub page.

What you are installing, in the docs’ own split

The docs give four ways in: Docker, pip, uv, and a desktop app from their GitHub. For a machine that should come back after a reboot, Docker with a restart policy is the boring choice. They distinguish a :main image, which carries embedding, speech, and reranking models so retrieval works immediately, from a :slim image when your chat model is already served elsewhere. Those bundled models are small helpers. They are not Qwen3.8.

The same page describes two faster-moving siblings. Open Terminal gives an agent a terminal and a file browser inside the chat. Open WebUI Computer is a separate harness for the real machine (files, git, a terminal), with approvals and a plan mode, installed with a uv tool, not by clicking chat. The docs say Computer moves faster and that pieces graduate into Open WebUI when they settle. If you want an always-on family chat box, install Open WebUI and Ollama and stop. If you want the agent to edit the disk, read the Computer page and leave approvals on. This guide does not walk through unattended root.

Their “connect an agent” list also names OpenClaw. This guide does not link a repository for it. Meta’s Glimmer post uses the same name for a scaffold, and there is still no confirmed project URL to publish.

Install it: copyable commands

Two paths, both from the projects' own documentation, read on 30 October 2026. The Mac path suits the Mac mini and the Mac Studio. The Linux path suits the GMKtec EVO-X2, which GMKtec lists with Windows 11 Pro or Ubuntu.

Mac: the Ollama app, then Open WebUI in Docker

Install Ollama by mounting ollama.dmg and dragging the app to Applications; Ollama's macOS page asks for macOS Sonoma (14) or newer. Then keep models loaded and allow parallel chats. Ollama's FAQ sets variables for the Mac app with launchctl and then restarts the app. For the three-person example below:

launchctl setenv OLLAMA_KEEP_ALIVE "-1"
launchctl setenv OLLAMA_NUM_PARALLEL "3"

Quit and reopen the Ollama app, then pull the model for your box. llama3.1:8b is the Llama 3 8B row of the worked example; qwen3:30b is Qwen3-30B-A3B.

ollama pull llama3.1:8b    # Mac mini M6 16 GB or Mac mini M5 Pro 24 GB
ollama pull qwen3:30b      # Mac Studio M5 Max 36 GB

Install Docker Desktop and start it. Then run the command from the Open WebUI README for Ollama on the same computer:

docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

That is the base-URL wiring: the container reaches Ollama through host.docker.internal, so nothing else is needed when both run on one Mac. Open the Mac's network address on port 3000 in a browser, for example http://192.168.1.100:3000 from another device, and create the first account, which Open WebUI makes the administrator. The Open WebUI quick start page also passes -e WEBUI_SECRET_KEY, set to the output of openssl rand -hex 32.

If Ollama runs on a different machine, the README drops the --add-host flag and sets OLLAMA_BASE_URL to that server instead. Put your Ollama server's address in place of the example:

docker run -d -p 3000:8080 -e OLLAMA_BASE_URL=http://192.168.1.60:11434 -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

That Ollama server has to accept connections from the network, which the Ollama FAQ sets with OLLAMA_HOST=0.0.0.0:11434. The FAQ describes no login for the Ollama API, so only do this on a network you trust.

Linux (the EVO-X2 on Ubuntu): the Ollama service, then Open WebUI in Docker

Install Ollama with its script, then open the service settings:

curl -fsSL https://ollama.com/install.sh | sh
sudo systemctl edit ollama.service

In the editor, add these lines under [Service], in the form the Ollama FAQ uses:

[Service]
Environment="OLLAMA_KEEP_ALIVE=-1"
Environment="OLLAMA_NUM_PARALLEL=3"

Reload, restart, set the service to start with the system as Ollama's Linux page does, and pull a model:

sudo systemctl daemon-reload
sudo systemctl restart ollama
sudo systemctl enable ollama
ollama pull qwen3:32b

Try the same docker run command as on the Mac. If Open WebUI cannot reach Ollama, the README's troubleshooting section explains why: inside the container, 127.0.0.1 points to the container, not to the host. Its fix runs the container on the host network, which moves Open WebUI to port 8080:

docker run -d --network=host -v open-webui:/app/backend/data -e OLLAMA_BASE_URL=http://127.0.0.1:11434 --name open-webui --restart always ghcr.io/open-webui/open-webui:main

Three boxes, three model tiers

Always-on means the weights stay loaded. The formula is params × bytes × 1.2 + 2. At Q4_K_M that is params × 0.6 + 2. The calculator is the place to check a context length this page does not cover.

Machine Leave loaded Do not expect
Mac mini M5 Pro 24 GB. K2-Horizon-7B at 6.2 GB all day. Qwen3.8-27B at 18.2 GB if the chat stays short. A 32B (21.2 GB) plus macOS plus a browser. A 70B.
Mac Studio M5 Max 36 GB. 27B with KV headroom. K2-32B. Muse Glimmer’s ~17 GB K-quant plus cache. 70B Q4, which is 44 GB. DeepSeek-V4-Flash.
GMKtec EVO-X2 128 GB. An aggressive quant of a large MoE (DeepSeek-V4-Flash is 284B total). Or a 70B Q4 at 44 GB with room. Discrete-GPU tokens per second. A second 70B resident beside the first.

The Strix Halo guide has the bandwidth and the Beelink caveat. The Apple Silicon guide has the configurations that were out of stock. There is no 64 GB Amazon ASIN here. Order that from Apple if you need it. Worked agent overhead, separate from this formula, is in the agent VRAM guide.

Worked example: three people on one always-on box

Always-on depends on two Ollama settings. Ollama's FAQ says a model stays in memory for 5 minutes after its last request and then unloads; a negative keep_alive, or OLLAMA_KEEP_ALIVE set to a negative value, keeps it loaded. It also says each model answers one request at a time by default (OLLAMA_NUM_PARALLEL=1), and that raising the number multiplies the context memory by the number of parallel requests.

For a household of three who may ask at the same moment, start Ollama with OLLAMA_KEEP_ALIVE=-1 and OLLAMA_NUM_PARALLEL=3, then size the box for three caches. Totals are the formula result plus three f16 caches of 8,192 tokens.

Model at Q4 Three chats at once Mac mini M6 16 GB (about 12 GB usable) Mac mini M5 Pro 24 GB (about 20 GB usable) Mac Studio M5 Max 36 GB (about 32 GB usable)
Llama 3 8B 10.0 GB Fits, 2.0 GB left Fits, 10.0 GB left Fits, 22.0 GB left
Qwen3-30B-A3B 22.7 GB No No Fits, 9.3 GB left
Qwen3-32B 27.6 GB No No Fits, 4.4 GB left

Not in those totals: the embedding, speech and reranking models the :main image bundles (their sizes are not on this site), and any browser or container, which our agent-workload guide pads at 2 to 4 GB. On a Mac both come out of the same pool. The 128 GB GMKtec EVO-X2 holds every row with room; the trade-off there is memory bandwidth, covered in the Strix Halo guide. Totals add the cache on top of the formula result, as our KV cache guide does, so they lean high by up to one 8K cache.

Keeping it on: power, restarts, remote access and logins

Idle power

Machine Vendor figure At idle, 30 days round the clock
Mac mini M6 16 GB Apple: 4 W idle, 70 W maximum, for the M6 with 16GB and a 256GB SSD 2.9 kWh
Mac mini M5 Pro 24 GB Apple: 6 W idle, 145 W maximum, published for the M5 Pro with 64GB and an 8TB SSD, not the 24 GB model 4.3 kWh
Mac Studio M5 Max 36 GB Apple: 7 W idle, 200 W maximum, for this configuration 5.0 kWh
GMKtec EVO-X2 128 GB GMKtec: no idle figure; the processor runs steadily at 120 W and peaks at 140 W under load, on a 230.1 W power adapter Not published

kWh = watts × 720 hours / 1,000. Apple's figures bracket a Mac: idle with only Finder open and default power settings, and maximum under load. A box answering three people sits between them, so a plug-in power meter gives your real number.

Restarts and power cuts

Remote access and logins

OpenBot (CopilotKit), if you want agents as channels

CopilotKit OpenBot (singular) is a self-hosted agent app. Its product page, read 2 October 2026, describes agents as channels and skills as slash commands, Docker Compose for a demo and Helm for production, conversations, vectors and permissions stored in your own Postgres, and no bundled model: an admin configures a provider. Agents that speak AG-UI, including LangGraph, Mastra and PydanticAI, can be registered as endpoints. It is not OpenBots.ai, a separate company in healthcare revenue-cycle automation.

For hardware, OpenBot is another front door to the same Ollama or vLLM server sized in the table above. What it adds is per-agent containers and signed-in browsers. Budget the 2 to 4 GB pad from our agent-workload guide for each one running at the same time, and on a Mac count it against the model's memory. Earlier notes called the repository alpha and that was not re-checked, so read the repository's own status before you treat it as a production deploy.

Which listing, after you know the tier

The Mac mini M6 16 GB is the smallest box in the worked example: three 8K chats on Llama 3 8B fit with 2.0 GB left of about 12 GB usable, and nothing larger in that table fits. Apple lists it at 4 W idle. Start with the Mac mini M5 Pro 24 GB if the household agent is a 7B to 14B that answers from a notes folder. Move to the Mac Studio when you want a 30B-class model to stay loaded for three people. Move to the EVO-X2 when the model itself does not fit in 36 GB. Ollama and Open WebUI run on all four; the weights decide which one you need.

Mac mini M6 16 GB on Amazon

ASIN B0HGGJZ92H. Amazon Associates link. The live price is on that page, not locked in here.

Mac mini M5 Pro 24 GB on Amazon

ASIN B0HGGHNQY6. Amazon Associates link. The live price is on that page, not locked in here.

Mac Studio M5 Max 36 GB on Amazon

ASIN B0HGKSQMX6. Amazon Associates link. The live price is on that page, not locked in here.

GMKtec EVO-X2 128 GB on Amazon

ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.

Sources

Questions

Does Open WebUI include a model?

No. You connect Ollama, vLLM, or another provider. The docs say a built-in path stays empty until you configure one. The :main Docker image does bundle embedding and speech models for retrieval and voice. That is not the chat model.

Which of the three Amazon boxes matches which model?

Mac mini M5 Pro 24 GB (B0HGGHNQY6): K2-Horizon-7B at Q4 (about 6.2 GB), and Qwen3.8-27B at Q4 (about 18.2 GB) only if you keep context short. Mac Studio M5 Max 36 GB (B0HGKSQMX6): that 27B with room, or K2-32B at about 21.2 GB. GMKtec EVO-X2 128 GB (B0F53MLYQ6): a large MoE such as an aggressive DeepSeek-V4-Flash quant. It is not CUDA.

Is Open WebUI Computer the same app?

No. The docs describe Computer as a faster agent runtime for the actual machine, installed separately, with approvals and plan mode. Open WebUI is the stable multi-user platform. You can point Open WebUI at a Computer workspace, but you do not need Computer to chat with Ollama.

Where do I ask a question about a fit?

Use the contact page on this site. Do not send hardware questions to a personal mailbox. Prices on Amazon move; the ASINs above are the listings, not a quote.

Will the box come back on its own after a power cut?

On a Mac, turn on "Start up automatically after a power failure" in System Settings, Energy. Add Ollama to Login Items, set Docker Desktop to start when you sign in, and choose an account under "Automatically log in as", which Apple makes unavailable when FileVault is on. The Open WebUI container then restarts with Docker because of --restart always.

How much power does the box draw when nobody is using it?

Apple lists the Mac mini M6 (16GB, 256GB SSD) at 4 W idle and the Mac Studio M5 Max 36 GB at 7 W idle, about 2.9 and 5.0 kWh in a 30-day month. GMKtec publishes no idle figure for the EVO-X2. Apple's idle figure assumes only Finder is open, so a plug-in power meter gives your real number.

Can I reach the box from my phone away from home?

Open WebUI's sharing guide recommends a private network such as Tailscale or Cloudflare Tunnels over public exposure, and HTTPS for remote access. Keep Ollama on its default 127.0.0.1 binding and let Open WebUI's accounts handle logins. Do not open ports 3000 or 11434 on your router.

A correction or a hardware question goes to the contact page.