Local models

A model running on your own computer costs nothing per question, has no usage limit, and keeps your questions at home. Ixel works with Ollama, LM Studio, and any model server that speaks OpenAI's chat API, on this computer or another of yours. Turn on Private, and only those models answer.

Picks from Artificial Analysis, read 4 October 2026

Set one up

Ixel doesn't run models itself: a model server does, and Ixel asks it. Pick one:

[agents.gpubox]
type = "http"
url = "http://gpu-box:8000/v1"   # the address the server shows; Ixel adds /chat/completions
model = "the-model-name"         # as the server lists it
label = "My GPU box (local)"

Which model fits your computer

The table shows the best open models a computer can run with each amount of memory. On a Mac, count all your memory, since the graphics share it. On a PC, count your graphics card's memory: a model that doesn't fit there still runs, but much more slowly.

The best open models overall: MiMo-V2.6-Pro (46, 1000B, about 602 GB), GLM-5.3-Flash (42, 320B, about 194 GB), K2 Horizon 375B A23B (31, 375B, about 227 GB). Those need a server, or a Mac with a lot of memory.

"Needs" is an estimate: about 0.6 GB for each billion parameters, which is a typical 4-bit download, plus 2 GB for the conversation. A smaller download of the same model needs less and answers a little worse; a larger one needs more. Model servers often name models differently from the benchmark, so search for the model's name in Ollama's or LM Studio's library.

Scores are on the same scale as the cloud picks, so you can see what you give up. Today's best local models score well below the best cloud ones, which is why the mix below has a strong cloud model check their work.

Models on your computers

In the app, Settings has a card for the models on your own computers:

• Look beside This computer checks each server's usual port on this computer. Beside Another computer, type its name, address or Tailscale name. Ixel looks only when you press Look, and only there: it never searches your network.
• Each server it finds lists its models. Add puts one on your panel, with no key. Models that can't answer a question, such as embedding models, are left out, and so are Ollama's cloud models (names ending in -cloud or :cloud), which run on ollama.com.
• A model you added shows its server's address under Models. Change the address to move the model to another of your computers that has it.
• Take off removes one of these models from your list. If it writes the verdict, verifies for Saver or decides for Triage, pick another model for that first. Ixel won't take off the last model on your panel, or the last one drafting for Saver.

Looking sends nothing of yours and no key: Ixel asks the server directly, never through a proxy, and doesn't follow a redirect. Ixel never gives a key to a server it finds or adds. (A server you add to the settings file by hand with token_env gets the key you set there.) Settings saves only an address on this computer or your own network, and asks the server there again before saving it. To add a company's model or a program, use ixel setup.

Get a model

Under each Ollama in Models on your computers, type a model's name as ollama.com/library shows it, such as qwen3:8b, or, for a Hugging Face model, hf.co/ followed by its name. Press Get, and that Ollama downloads it onto the computer it runs on, showing how far along it is. Then press Add beside it.

• Where it comes from: Ixel only asks your Ollama to get the model. Ollama downloads it from ollama.com, or from Hugging Face for an hf.co/ name, as ollama pull does. Those sites see that computer's address and the model's name. Ixel sends nothing else.
• One download at a time. Closing the page stops the download. Getting the same model again carries on from where it stopped.
• Cloud models aren't offered. A name ending in -cloud or :cloud runs on ollama.com, not on your computer.
• LM Studio and the other servers download models in their own apps.

Use another computer

A model server on another of your computers, such as a desktop with a big graphics card or a Mac with a lot of memory, works the same way. In Settings, type its name, address or Tailscale name beside Another computer (mac-mini, 192.168.1.20 or mac-mini.tail1234.ts.net) and press Look. ixel setup asks for one too. With no port, Ixel tries each server's usual port there.

1. Let other computers in

   On that computer, in LM Studio, turn on Serve on Local Network in its Developer tab. For Ollama, set OLLAMA_HOST=0.0.0.0 and restart it. That computer's firewall also has to allow the port.

   Anyone on that network can use it

   Out of the box, Ollama and LM Studio don't ask for a password. Once a server lets other computers in, anyone on the same network can ask it questions, and with Ollama, download or delete its models. Do this only on a computer that stays on your home network, never on a laptop you use on café or office Wi-Fi. The setting stays on wherever the laptop goes.

2. Use a name Ixel can tell is yours

   Ixel adds only a computer on your own network or your Tailscale network: a private address, a Tailscale address, or a name that leads only to those. Ixel checks the name again each time it asks a question, so the name has to look like yours: a short name like mac-mini, or one ending in .local, .lan, .internal, .home.arpa or .ts.net. Your router's own names, like mac-mini.fritz.box, don't count; use the address instead. A name you type is looked up like any computer name, through your usual DNS.

A server anywhere else, such as one on the internet, goes in your settings file by hand and isn't counted as yours. Calls to your own computers are free in the cost line. Plain http:// across your network isn't encrypted; over Tailscale it is.

Private

Turn on Private in Settings, under Asking, and only the models on your own computers answer. Or set it in your settings file:

[review]
private = true

A model counts as yours when Ixel asks it directly, with no key, at an address on this computer or your own network. These sit out:

• Every company's model, and every program Ixel starts (Claude Code, Codex, OpenCode and the rest), since Ixel can't see where a program sends things.
• A verdict writer that isn't one of yours: the author of the best-rated answer writes the verdict instead.
• Saver's verifier, if it isn't one of yours. Saver won't run until you pick one of yours to verify.
• Triage, unless one of your models decides.
• A server that takes a key or is marked billing = "api": that's most likely a gateway that passes questions on to a company.
• Ollama's cloud models (-cloud or :cloud), which run on ollama.com.

Private can't see what a server does with a question. A server on your network still counts as yours even if it passes questions on to a company. That includes a proxy like LiteLLM set up without a key, or an app connected to a cloud model. Add only servers that run the model themselves.

With Private on:

• If none of your models can answer, Ixel says so and sends nothing.
• Sound isn't sent to OpenAI or Groq to be written out. The only exception is a speech server on this computer that you set as openai_url or groq_url under [sound] in the settings file. That server still needs your OpenAI or Groq key saved, and Ixel sends it that key.
• The terminal chat, ixel review and ixel ask use only your models, and /handoff in Ask is off, since Handoff's coding agents send work to their companies. An ixel chat that was already open keeps the setting it started with: close it and start it again after you turn Private on.
• A conversation asked with Private on isn't carried into a follow-up once Private is off.

Mix local and cloud

For a cheaper panel, let local models do the drafting and reviewing, and have one strong cloud model check their work. That's Saver:

[saver]
verifier = "opus"                 # one strong model, on a key or a subscription
drafters = ["qwen", "gpubox"]     # your local models
verifier_effort = "low"           # it only has to check

Or set it up in the app under Settings, Saver. The drafters' calls are free. The big model reads their drafts and replies in a few words, so it costs far less than writing the answer itself. The four jobs says more. With Private on, the verifier has to be one of your own models too.

Each review's cost line counts local calls as free:

  Cost: $0.042 on API keys · 3 calls on your subscriptions · 1 free on your own hardware · 12.3k tokens in, 2,100 out

Pictures

A local model doesn't get the pictures you attach in the app, since most can't read them. For one that can, such as a vision model, turn on pictures for it in Settings, Pictures, or add accepts = ["image"] to its section of the settings file.

Thinking settings

A model server on your computer gets thinking settings from minimal to high, so xhigh and max are sent as high. Whether the model uses the setting is up to the model and the server. The table above counts only the settings Ixel can send.

What stays on your computer

With Private on, your questions, your code and the answers go no further than your own computers, as long as your servers run the models themselves. A few parts of Ixel reach the internet for other reasons, and none of them sends your questions:

• The terminal chat checks GitHub for a new version once a day. [updates] check = false turns it off.
• Get a model has your Ollama download from ollama.com or Hugging Face when you press Get.
• Making pictures sends your description to xAI or OpenAI, but only when you use it. With Private off, Sound sends a recording to OpenAI or Groq the same way.
• The Board reads a project's pull requests from its git host when you open it.
• Opening Settings asks each company you've saved a key for which models it has, and Health's Check now, ixel status and ixel doctor --check ask each model whether it answers. These send your key, never a question, and Private doesn't stop them.

Privacy has the whole list.