LM Studio Setup — Skales Guide

1 · Download LM Studio

1

Go to lmstudio.ai and download the installer for your OS (Windows, macOS, or Linux).

2

Install and launch LM Studio. The app opens to the Discover tab where you can search for models.

💡 Recommended Models

For a great first experience, search for: Llama-3.1-8B-Instruct, Mistral-7B-Instruct, or Qwen2-7B-Instruct. Choose a Q4_K_M quantization for a balance of speed and quality.

2 · Download a Model

1

In LM Studio, click the Search (🔍) icon in the left sidebar to open the model browser.

2

Search for a model name (e.g. llama-3.1-8b-instruct), select a version, and click Download.

3

Wait for the download to complete. Models are stored in ~/LM Studio/models/.

3 · Start the Local Server

1

Click the Local Server icon (↔) in the left sidebar.

2

Select your downloaded model from the model dropdown at the top.

3

Click Start Server. You'll see a green status: Server running on port 1234.

ℹ️ Default Port

LM Studio's default server port is 1234. You can change this in the server settings if another app is already using that port.

4 · Configure Skales

1

Open Skales → Settings → AI Providers. The page opens on the provider grid: one tile per provider, with badges for what each one can do.

2

Find the LM Studio tile and switch it on. LM Studio has its own tile — you no longer go through Custom (OpenAI-compatible) for it.

3

Open the card. The endpoint is pre-filled with LM Studio's default:

http://127.0.0.1:1234

Change it only if you moved the server to another port or another machine. Use 127.0.0.1 rather than localhost — it avoids IPv6 resolution issues.

4

There is no API Key field to fill. LM Studio is a keyless provider in Skales, the same way Ollama is: the card asks for an address, not a credential.

5

Click Fetch Models. Skales asks your running LM Studio server what it has loaded, and the model appears in the picker. LM Studio ships no built-in model list in Skales — everything you see here comes from your own server.

6

Pick your model and start chatting. Tool calling is on by default; how well it works depends on the model, not on Skales.

✅ Verification

Click Test Connection on the card. The connection map checks six links in order: address, server, model, context, tools and a real test turn. It turns green only when the test turn comes back, and it names the first link that breaks together with what to do about it. The same map opens from a chat error with Check connection, and the diagnostics report carries it too.

LM Studio's server log shows each incoming request as well.

Tips & Best Practices

GPU Acceleration LM Studio automatically detects and uses your GPU. Check the server settings to configure GPU layer offloading.

Multiple Models You can have multiple models downloaded. Switch between them in LM Studio's server tab — then click Fetch Models in Skales to update.

Context Size Skales needs a context length of at least 12,288 tokens, 16,384 or more is comfortable. LM Studio often loads models at 4,096, which is too small for the Skales prompt: raise Context Length in the model settings and load the model again.

Use 127.0.0.1 Always use 127.0.0.1 in the endpoint URL, not localhost — avoids IPv6 issues on Windows.

Quantization Q4_K_M is the sweet spot. Download Q5_K_M if you have extra RAM. Q2_K if you're on a very constrained system.

Keep Server Running LM Studio's server must be running while you use Skales. It doesn't start automatically — open LM Studio and start the server each session.

Troubleshooting

Problem	Fix
Connection refused	Start the local server in LM Studio first. Click the ↔ icon → select a model → click Start Server.
Fetch Models returns empty	The server is not running, or the port doesn't match. Verify LM Studio shows "Server running on port 1234" and check your Skales endpoint URL.
Slow responses	Try a smaller model (3B–7B), or enable more GPU layers in LM Studio's server settings under GPU Offload.
"Ollama not installed"	You are on the wrong tile. Switch on LM Studio in Settings → AI Providers — not Ollama, and not Custom (OpenAI-compatible).
Tool calls failing	Click Test Connection: the Tools link says whether the model accepts tool calls. If it does not, Skales still chats with this model, without tools. For files, the browser and the other tools, load a model with tool support (LM Studio marks them with a hammer icon). Use an instruct-tuned model, not a base model.
The model picker stays empty	LM Studio contributes no built-in model list — the picker fills from your server. Load a model in LM Studio, keep the server running, then click Fetch Models again.
Wrong port	Check LM Studio's server tab for the actual port in use. Update the endpoint URL in Skales Settings accordingly.
The card is green but the chat fails	Click Test Connection and read the connection map: the first red link is where it breaks. A list of models alone never turns the map green, only a test turn that comes back does.
No model loaded	Load a model in LM Studio, then check again. LM Studio: My Models, Load.
Context too small	Raise the context length to at least 16,384 tokens. LM Studio: model settings, Context Length, then load the model again.
This model does not accept tool calls	Skales still chats with this model, without tools. For files, the browser and the other tools, load a model with tool support (LM Studio marks them with a hammer icon).

LM Studio / Skales on Windows with WSL2

WSL2 is a small virtual machine with a network of its own. localhost on one side is not the other side, so a server that answers perfectly on Windows can be unreachable from WSL, and the other way round. The connection map on the LM Studio card names which of the cases below you are in.

Which side listens where

Setup	Address to use in Skales
Skales and LM Studio both on Windows	`http://127.0.0.1:1234/v1`. Nothing else to do.
Skales on Windows, the server (LM Studio, Ollama, llama.cpp) inside WSL2	If the server runs inside WSL2, it listens inside WSL, not on Windows. Use the address that `hostname -I` prints in WSL, or turn on mirrored networking (`networkingMode=mirrored` in `.wslconfig`) and restart WSL. Start the server so it listens on the network (`OLLAMA_HOST=0.0.0.0`, llama.cpp `--host 0.0.0.0`).
Skales inside WSL2, LM Studio on Windows	Skales runs inside WSL2, where localhost is WSL itself, not Windows. Use the Windows address that `ip route show default` prints after `via`, and switch on Serve on Local Network in LM Studio.
LM Studio on another computer in your network	The address of that computer, for example `http://192.168.1.20:1234/v1`, with Serve on Local Network switched on.

Checklist

1

Serve on Local Network. LM Studio answers other computers only with Serve on Local Network switched on (Developer tab, server settings). From WSL, Windows counts as another computer.

2

Firewall. Allow port 1234 (or the port you set) through the Windows firewall of the computer that runs the server. A request that is dropped by the firewall shows up as "did not answer within ...s" on the server link.

3

Model loaded. Load a model in LM Studio, then check again. LM Studio: My Models, Load.

4

Context length. Raise the context length to at least 16,384 tokens. LM Studio: model settings, Context Length, then load the model again.

5

CORS. CORS settings do not matter here: Skales talks to the server directly, not from a web page.

6

Check it. Click Test Connection on the LM Studio card. When the map is green all the way to Test turn, the chat works too.

LM Studio + Skales