Run Gemma 4 in LM Studio: 12B Is Here for 16GB Macs

Not everyone wants to type commands into a terminal to run a local AI model. If you'd rather have a proper desktop app with a search bar, a chat window, and a one-click download button, LM Studio is what you're looking for.

This guide walks you through setting up LM Studio and getting Gemma 4 running on your machine — from installation to your first conversation, plus how to use the built-in API server for development.

What Is LM Studio?

LM Studio is a free desktop application for running large language models locally on macOS, Windows, and Linux. Think of it as a complete local AI workstation with a graphical interface.

What sets it apart from CLI tools like Ollama:

• Built-in model browser — search, discover, and download models without leaving the app
• Chat interface — a clean UI for conversations, similar to ChatGPT
• Hardware-aware loading — LM Studio detects your GPU and RAM, then suggests the best quantization and settings automatically
• OpenAI-compatible API — serves models as a local API that works with any tool expecting the OpenAI format
• Supports GGUF and MLX formats — GGUF for all platforms, MLX for optimized Apple Silicon performance

LM Studio supports Gemma models in both formats, so whether you're on a MacBook, a Windows desktop, or a Linux workstation, it has you covered.

Step 1: Install LM Studio

Download the installer for your OS from lmstudio.ai/download:

• macOS: Download the .dmg, drag to Applications, and launch
• Windows: Run the .exe installer
• Linux: Follow the instructions on the download page

Once installed, open LM Studio. You'll see the main interface with a model search bar, a chat panel, and a sidebar for settings.

Step 2: Download Gemma 4

There are two ways to get Gemma 4 into LM Studio.

Option A: In-App Model Browser (Easiest)

1. Press ⌘ + Shift + M (Mac) or Ctrl + Shift + M (Windows/Linux) to open the model search
2. Type "Gemma 4" in the search bar
3. Browse the results — LM Studio will highlight which variants are compatible with your hardware
4. Pick a variant and click Download

LM Studio automatically suggests the best quantization level for your machine. If you have 16 GB of RAM, it'll now steer you toward the new Gemma 4 12B (released June 3, 2026) rather than E4B — it's the model LM Studio's core 16 GB-MacBook audience should reach for first. If you have 32 GB+, it may suggest the 26B MoE.

Option B: From Hugging Face

If you prefer to browse models on Hugging Face first:

1. Enable LM Studio in your Hugging Face Local Apps Settings
2. Go to any Gemma 4 model page on Hugging Face
3. Click the "Use this model" dropdown and select LM Studio
4. The model downloads and appears in LM Studio automatically

This is handy if you've found a specific community-quantized version or a fine-tuned variant you want to try.

The New 12B: Built for 16 GB MacBooks

Google added a fifth Gemma 4 size on June 3, 2026, and it's the one most LM Studio users have been waiting for. Gemma 4 12B is a dense model (no MoE) with a unified, encoder-free multimodal design — it takes text, images, and audio as input natively, and it keeps the family's 256K context window. At Q4 quantization the weights are about 6.7 GB, which is exactly why it matters: it's the first model in the lineup that delivers real workstation-adjacent quality while comfortably fitting a 16 GB unified-memory Mac or a 16 GB VRAM GPU, closing the gap between the E4B edge model and the 24 GB+ territory the 26B/31B models need.

In LM Studio, search for "Gemma 4 12B" in the model browser (⌘/Ctrl + Shift + M) and you'll find it in both formats:

• GGUF — lmstudio-community/gemma-4-12B-it-GGUF, available in Q4_K_M, Q6_K, and Q8_0 quantizations. There's also a QAT (quantization-aware training) build, lmstudio-community/gemma-4-12B-it-QAT-GGUF, which trades a slightly larger download for better quality retention at low bit depths.
• MLX — lmstudio-community/gemma-4-12B-it-MLX-4bit for Apple Silicon, tuned to run through Metal instead of the GGUF/llama.cpp path.

If you're on a 16 GB MacBook Air or Pro, this is the model to try first — it's a meaningful step up from E4B without needing to upgrade hardware.

Which Model to Download?

If you're not sure, here's a quick reference:

Step 3: Start Chatting

Once the download finishes:

1. Click on the model name in your model list to load it
2. A progress bar shows the model loading into memory
3. When it's ready, the chat window activates — type a message and hit Enter

That's it. No configuration files, no terminal commands, no environment variables. You're chatting with Gemma 4 locally.

A few things to try in your first session:

• "Explain how a car engine works, like I'm 10 years old." — Tests general knowledge and tone control
• "Write a Python function that finds duplicate files in a directory." — Tests code generation
• "What are the pros and cons of TypeScript vs. JavaScript?" — Tests structured reasoning

Working with Images

Gemma 4 is multimodal — it understands images natively. In LM Studio's chat interface, you can drag and drop an image directly into the message box, or use the attachment button to browse for a file.

Try these:

• Drop in a screenshot and ask "What does this UI show?"
• Attach a photo of a receipt and ask "Extract the total and date"
• Paste a chart and ask "What trend does this show?"

The model processes the image entirely on your machine — nothing is uploaded anywhere.

A note on audio: Gemma 4 12B natively understands audio at the model level — no separate speech-to-text step required. As of this writing, LM Studio's chat interface only supports attaching images, not audio files, so you can't yet drag in a voice memo the same way you drag in a photo. If you need audio input today, look at running the model directly through Hugging Face transformers or mlx-lm instead. Keep an eye on LM Studio's release notes — audio attachment is a natural next step given the model's capabilities.

Using the Local API

LM Studio can serve your loaded model as a local API server, which is useful for integrating Gemma 4 into your own apps, scripts, or development tools.

Starting the Server (GUI)

1. Go to the Developer tab in LM Studio
2. Load a model by pressing ⌘/Ctrl + L and selecting Gemma 4 from your downloaded list
3. The server starts automatically on http://localhost:1234

Starting the Server (CLI)

If you prefer the terminal, LM Studio includes a CLI called lms:

# List your downloaded models
lms ls

# Load a model
lms load <model_key>

# Start the API server
lms server start

Making API Calls

The API is OpenAI-compatible, so any code or tool that works with the OpenAI API works with LM Studio — just change the base URL:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio",
)

response = client.chat.completions.create(
    model="gemma-4-e4b",
    messages=[{"role": "user", "content": "What is the capital of France?"}],
)

print(response.choices[0].message.content)

# Or with curl
curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-e4b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Swap the model field for whichever variant you've loaded — e.g. google/gemma-4-12b for the new 12B GGUF build, or the MLX identifier if you loaded that instead.

LM Studio also has its own native SDKs — lmstudio-python and lmstudio-js — that offer additional features like stateful chats and model management (download, load, unload) through code.

Tool Use and Function Calling

LM Studio supports Gemma 4's native function calling capabilities through the API. You can define tools in the standard OpenAI format and the model will generate structured function calls:

response = client.chat.completions.create(
    model="gemma-4-e4b",
    messages=[{"role": "user", "content": "What's the weather in Tokyo?"}],
    tools=[{
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a city",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string"}
                },
                "required": ["city"]
            }
        }
    }]
)

This enables building agentic workflows where Gemma 4 decides when to call external tools and how to chain results together — all served from your local machine.

LM Studio vs. Ollama: Which Should You Use?

Both are excellent tools for running Gemma 4 locally. The choice depends on your workflow:

Use LM Studio if you want a polished desktop experience, prefer browsing models visually, or are new to local AI.

Use Ollama if you're comfortable with the terminal, want to script model management, or are running a headless server.

Both serve an OpenAI-compatible API, so your downstream code works with either one — switching between them is just a URL change.

Tips for a Better Experience

Let LM Studio choose your settings first. It detects your hardware and picks reasonable defaults for GPU layers, context length, and thread count. Only tweak these after you've confirmed the model runs.

Start with a smaller context length. The default might be set to the model's maximum (128K or 256K tokens). If you're just chatting, reducing this to 4K–8K frees up significant memory and speeds up loading.

Use MLX on Apple Silicon. If you're on a Mac with an M-series chip, look for MLX-format models — they're optimized for Apple's GPU architecture and typically run faster than GGUF on the same hardware.

Import custom models with the CLI. If you've converted or fine-tuned a Gemma 4 model to GGUF format yourself, use lms import to add it to LM Studio. It will auto-detect the model architecture and make it available in the app.

What's Next?

You've got Gemma 4 running in LM Studio with a chat UI and a local API. A few directions from here:

• Compare models side by side — download multiple Gemma 4 variants and switch between them to see how quality scales with size
• Connect it to your tools — the OpenAI-compatible API means VS Code extensions, Obsidian plugins, and other apps can use your local Gemma 4 as a backend
• Check our other guides — Hardware Requirements if you want to optimize, or Benchmarks to understand what each variant is capable of

Running AI locally shouldn't feel like a systems administration task. With LM Studio and Gemma 4, it doesn't have to.