QUICKSTART.md

QUICKSTART: Get Running Fast

Goal: Get a local inference engine running, a model loaded, and a coding interface connected — in under 10 minutes.

This guide follows the reference stack this documentation was built around. It's one validated path, not the only one. The Reading List covers the broader landscape of inference engines, models, and interfaces if you want to explore alternatives before or after.

For: 16GB/32GB Windows 11 systems (Ryzen AI, Intel Lunar Lake preferred).

16GB users: May work on a case-by-case basis (HW limitations). Please pay attention to the info in the sidebars.

All commands run in Windows Terminal (PowerShell).

---

Prerequisites

• Windows 11 (recent build, 24H2 or later recommended)
• 32GB RAM (16GB RAM supported with limitations, check sidebars)
• Updated GPU drivers (Ryzen AI / Intel Lunar Lake / compatible)
• Administrator access for winget installation

---

Step 1: Install LM Studio

winget install ElementLabs.LMStudio --accept-package-agreements --accept-source-agreements

Verify installation:

lms --version

Expected: Version string displays. If command not found, restart terminal or add LM Studio to PATH.

Note: The lms command-line interface provides all model management. The GUI is available but not required for this setup.

---

Step 2: Start the LM Studio Server

lms server start

Verify server running:

lms server status

Expected: Server running, listening on a port (typically 1234).

Troubleshooting: If port conflict occurs, LM Studio docs show how to configure alternate ports.

---

Step 3: Download the Model

The reference configuration for 32GB systems uses Qwen 3.6 35B with MoE architecture and q4_k_m quantization:

lms get qwen/qwen-3.6-35b-instruct-gguf --quant q4_k_m

Verification note: The exact model identifier may vary. Check available models with lms search qwen. Use the Qwen 3.6 35B variant with q4_k_m quantization.

16GB sidebar: For 16GB systems, use a 9B model instead:
lms get qwen/qwen-2.5-9b-instruct-gguf --quant q4_k_m

Download takes time. The model is ~18-20GB. Progress displays in terminal.

This is a good time to grab coffee. Model downloads take a few minutes.

---

Step 4: Load Model with Stability Settings

lms load qwen/qwen-3.6-35b-instruct-gguf --context-length 32768 --parallel 1

Why these settings:

• --context-length 32768: Tested context window size that works reliably on 32GB with other applications running
• --parallel 1: Single request at a time prevents memory contention

Verify model loaded:

lms ps

Expected: One model showing "loaded" status. Memory usage visible.

16GB sidebar: Same command, just reference the 9B model name instead.

---

Step 5: Install opencode

# Verify you have node/npm first
node --version

# Install opencode globally
npm install -g @anthropic-ai/opencode

Don't have Node.js? Download from nodejs.org (LTS version recommended).

Developer note: If you already have opencode installed, verify version is current: opencode --version

Verify installation:

opencode --version

---

Step 6: Connect opencode to LM Studio

Create opencode.json in your project directory (or use global config):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "lmstudio": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LM Studio",
      "options": {
        "baseURL": "http://localhost:1234/v1"
      },
      "models": {
        "qwen": {}
      }
    }
  },
  "model": "lmstudio/qwen"
}

Configuration note: These keys match opencode's documented schema. For full schema, see CONFIG.md.

Test the connection:

opencode "write a hello world PowerShell script"

Expected: opencode connects to LM Studio, sends request, returns generated script.

Success looks like: Model responds with PowerShell code. Response time varies (10-30 seconds typical for first request).

Tip

Field note: Treat a successful one-shot request as a connection check, not proof that an iterative IDE agent will be reliable. In engagement testing, one-shot local generation worked, while iterative agentic workflows timed out across the tested configurations. For longer back-and-forth work, use the local chat interface and keep the task focused. See the assessment findings for scope and context.

---

Step 7: Verify Single-Model Setup

lms ps

Confirm: Only one model shows "loaded" status. This is important for 32GB stability.

Why it matters: Loading multiple large models simultaneously exhausts RAM and causes system paging. Single-model operation keeps inference fast and stable.

Tip

Field note: If the system slows down or freezes, reduce competing load before assuming the model is broken: close background applications, keep prompts narrow, and clear or restart long sessions. These were practical participant workarounds on constrained devices. See the hardware findings for context.

---

Next Steps

It works? Great. Now:

1. Understand what you built: Read SETUP.md for how LM Studio + opencode work together
2. See it in action: Check USE_CASES.md for PowerShell generation and troubleshooting examples
3. Customize: Read CONFIG.md when you want to tune settings or switch models
4. Reality check: Read CAVEATS.md for honest assessment of what you're getting—tradeoffs, costs, and limitations
5. Start the challenge: Read CHALLENGES.md for the weekly challenge program — your setup just completed Week 1's prerequisite

Having issues? See NOTES.md troubleshooting section for common problems and solutions.

---

Ollama (Future)

Ollama integration is planned but not documented yet. Focus on LM Studio for now—it's tested and works reliably for this configuration. Ollama guidance will be added when we can provide the same quality of verified instructions.

---

You're set up. Time to put it to work.

See Use Cases → | Understand the Setup →

---