Models that never leave your machines.
Run open models on your own hardware. Download one, and TurnPanel puts it on the machine best able to run it, loads it when a job needs it, and makes it available in every space.
UnmeteredAcross every machineWorks air-gapped
- Qwen3 32BKept loadedgpu-boxLoaded21.4 GB
- Gemma 3 12Bstudio-macLoaded8.1 GB
- Llama 3.3 70Bgpu-boxOn disk42.5 GB
- Qwen3 8Bstudio-macLoading 64%5.1 GB
gpu-box37.4 of 64 GB
studio-mac21.2 of 36 GB
Models Spaces
From download to answer, handled.
You choose the model. TurnPanel looks after where it lives, when it loads and how every space reaches it.
- 01
Pick a model
From the catalogue, by its Hugging Face id, or by importing a file you already have. Downloads resume after interruptions and are checked against the repository’s SHA-256.
- 02
TurnPanel places it
TurnPanel benchmarks each machine’s memory bandwidth and model speed, and suggests the one best able to run the model. Move it to another machine any time.
- 03
It loads on demand
llama.cpp’s server runs the model, on the machine’s GPU. The first request waits while it loads, and the job shows the progress.
- 04
Every space can use it
Inside each space, one OpenAI-compatible endpoint lists every model in the workspace. pi uses it, and so can the apps and scripts your agents build.
Your memory, planned for you.
Each machine has a planner that decides what stays loaded, so switching models is quick and nothing runs out of room halfway through a job.
Never loads what can’t fit
“Qwen 35B needs 38 GB; 22 GB is free on studio-mac.” A clear reason instead of a machine swapping itself to a halt.
One budget with your spaces
Loaded models and running spaces share each machine’s memory, so neither starves the other.
Least used goes first
When a model needs room, the one used longest ago unloads. It stays on disk and loads again when a job needs it.
Keep the important ones loaded
Pin a model to a machine and it never unloads. Eject any other to free its memory at once.
Fair between spaces
Loads from different spaces take turns, so one busy space can’t hold up everyone else.
Embeddings built in
A small embedding model runs on every machine for memory search, always loaded and out of your way.
One GPU box. Every space can use it.
With TurnPanel Server, machines join into one workspace. A model on the GPU box is there for a space running on any other machine; requests go to wherever the model lives.
If that machine goes offline, TurnPanel uses the same model on another machine if one has it, then the space’s fallback model, and otherwise tells the job which machine is missing.
How machines join- gpu-box
- Qwen3 32B, Llama 3.3 70B. Serves every space.
- studio-mac
- Gemma 3 12B and Qwen3 8B, on the Mac’s GPU.
- front-door
- No models. Its spaces still use both machines’ models.
- If one is offline
- The same model elsewhere, then the space’s fallback model.
Nothing leaves. Not even offline.
Local models run on your machines, outside the spaces’ sandboxes and reachable only from inside your workspace.
- Air-gapped
Air-gapped mode stops every outside call. Import models from files, and everything keeps working.
- No keys in spaces
Spaces reach models through TurnPanel’s endpoint. Provider keys stay with the server, never inside a space.
- Unmetered
No tokens to count and no bill at the end of the month. Your hardware is the only limit.
- Mix when you want
Add your own providers, or TurnPanel’s hosted models with Pro, beside local ones in the same picker.
pi, your apps, your scripts.
pi, the open-source agent built into TurnPanel, works with any model you add, local ones included. So do the apps and scripts agents build, through the endpoint in every space.
Claude Code, Codex, Cursor and the other agents use their makers’ own models, as they would on your laptop.
| On a Mac | Apple silicon with at least 5 GB of free memory. |
|---|---|
| On a PC or server | A dedicated graphics card with at least 6 GB of video memory. |
| Formats | GGUF models, including split files and vision projectors. |
| Context | Loaded with up to 16K tokens of context. |
Your models, on your terms.
Free on your own hardware, in TurnPanel Personal and Server alike.
Free on your own hardwareOpen source