🤖 Ghostwritten by Claude Opus 4.6 · Curated by Tom Hundley
OpenClaw just dropped v2026.3.7-beta.1 about 18 hours ago, and it has two features I'm genuinely excited about: a new Docker slim variant that shrinks your OpenClaw container down to a fraction of its normal size, and first-class Google Gemini 3.1 Flash-Lite support that lets you analyze images, audio, and video right from your OpenClaw setup. If you've been running OpenClaw on a small server (or paying for cloud hosting by the gigabyte), the slim variant alone could save you real money. And if you've wanted a fast, affordable AI model that can look at things — screenshots, PDFs, photos — Gemini 3.1 Flash-Lite is now plug-and-play. Let me walk you through both features step by step, no coding experience required.
TL;DR: This release delivers a smaller, faster Docker image and a new budget-friendly AI model that understands images and media — both configured with just a couple of settings.
Contributor @sallyom submitted this via PR #38479, and it touches two very different parts of OpenClaw:
Docker multi-stage build restructuring — Think of this like packing a suitcase. The old way threw everything in: build tools, temporary files, the kitchen sink. The new slim variant packs only what OpenClaw actually needs to run. Smaller suitcase, same vacation.
Google Gemini 3.1 Flash-Lite integration — This is a new AI model from Google that's designed to be fast and cheap while still being able to understand images, audio, and video. OpenClaw now treats it as a first-class citizen with proper aliases, model normalization, and CLI support.
Let's set them up.
TL;DR: Add OPENCLAW_VARIANT=slim to your Docker setup to get a minimal runtime image that uses significantly less disk space and memory.
Docker is like a shipping container for software. Instead of installing OpenClaw directly on your computer (which can get messy), you run it inside a container that has everything it needs. If something breaks, you throw away the container and start a new one. Clean and simple.
The standard OpenClaw Docker image includes everything needed to build and run the software. The slim variant strips out all the build-time tools — compilers, package managers, development libraries — and keeps only what's needed to run. Here's what that means in practice:
| Feature | Standard Image | Slim Variant |
|---|---|---|
| Includes build tools | Yes | No |
| Relative image size | Larger | Significantly smaller |
| Startup speed | Normal | Faster |
| Best for | Development / contributing | Running OpenClaw day-to-day |
| Security surface area | Larger (more installed software) | Smaller (less attack surface) |
If you're already running OpenClaw with Docker, here's how to switch. Open the file where you configure your OpenClaw Docker setup — this is usually a file called docker-compose.yml or similar. Look for the section that defines your OpenClaw service and add this environment variable:
environment:
- OPENCLAW_VARIANT=slimIf you're starting fresh, your basic Docker Compose file would look something like:
services:
openclaw:
image: openclaw/openclaw:v2026.3.7-beta.1
environment:
- OPENCLAW_VARIANT=slim
ports:
- "3000:3000"Then restart your container. If you're using Docker Compose, that means running this command in your terminal:
docker compose up -d --pull alwaysThat --pull always flag ensures you're grabbing the latest v2026.3.7-beta.1 image with slim support baked in.
TL;DR: Gemini 3.1 Flash-Lite is a fast, affordable AI model that understands images and media — and OpenClaw now supports it with simple model aliases and full CLI compatibility.
You might already be using models like GPT-4o or Claude in OpenClaw. Gemini 3.1 Flash-Lite fills a different niche:
| Model | Speed | Cost | Media Understanding | Best For |
|---|---|---|---|---|
| GPT-4o | Moderate | Higher | Yes (images) | Complex reasoning |
| Claude Sonnet | Moderate | Higher | Yes (images, docs) | Nuanced writing |
| Gemini 3.1 Flash-Lite | Very fast | Lower | Yes (images, audio, video) | Quick tasks, media analysis, budget use |
The key advantage: Gemini 3.1 Flash-Lite handles images, audio, and video while being designed for speed and lower API costs. If you want to ask your AI to "look at this screenshot and tell me what's wrong," this model is built for exactly that.
Before OpenClaw can talk to Gemini, you need an API key — think of it like a password that lets OpenClaw access Google's AI. Here's how to get one:
Important: Treat this key like a password. Don't share it, don't paste it in public chats, don't commit it to a public code repository.
Add your API key to your OpenClaw configuration. In your docker-compose.yml, add it as an environment variable:
environment:
- OPENCLAW_VARIANT=slim
- GOOGLE_API_KEY=your-api-key-goes-hereOpenClaw v2026.3.7-beta.1 includes model normalization and aliases for Gemini, which means you don't need to remember the exact technical model name. You can reference it as:
gemini-3.1-flash-litegoogle/gemini-3.1-flash-lite-previewBoth will work. OpenClaw maps them to the right place automatically.
To set it as your default model, add this to your OpenClaw config:
- DEFAULT_MODEL=gemini-3.1-flash-liteRestart your container with docker compose up -d and you're live.
TL;DR: You can send images directly to Gemini Flash-Lite through OpenClaw's CLI or chat interface and get instant analysis — perfect for debugging screenshots or reviewing documents.
One of the best features of the Gemini integration is media understanding — the ability to send an image (or audio, or video) and ask questions about it. Here's a prompt you can try right now:
In your OpenClaw CLI, try:
/model gemini-3.1-flash-liteThis switches your active model. Then send an image with a question like:
"Here's a screenshot of my app's error page. What's the error message saying, and what should I try to fix it?"
Or for non-coding use:
"I'm attaching a photo of this form I received in the mail. Can you read it and summarize what it's asking me to do?"
The model will analyze the visual content and respond in plain English. According to Google's own benchmarking, Flash-Lite models are optimized for low-latency responses, meaning you should get answers back noticeably faster than with larger models.
Since we're adding a Google API key today, let's talk security for a moment. According to the GitGuardian 2024 State of Secrets Sprawl report, over 12.8 million new secret credentials were exposed in public GitHub repositories in a single year. API keys are the number one culprit.
Here are three rules to live by:
For a deeper dive on keeping your credentials safe, check out our earlier posts in the Vibe Coder Security series.
No. The slim variant is selected through an environment variable (OPENCLAW_VARIANT=slim), not a separate installation. You add it to your existing Docker configuration and restart your container. OpenClaw handles the rest during the image build process.
Absolutely. OpenClaw supports multiple AI models simultaneously. You can switch between them using the /model command in the CLI or by setting per-conversation model preferences. The DEFAULT_MODEL setting just controls which one is used when you don't specify.
Google AI Studio provides a free tier with usage limits that's generous enough for personal use and experimentation. Beyond that, Gemini Flash-Lite is designed as a budget-friendly option — the "Lite" designation specifically targets cost-conscious use cases. Check Google's current pricing page for exact rates, as these change frequently.
Gemini 3.1 Flash-Lite supports images, audio, and video understanding. In practice through OpenClaw, image analysis is the most common use case — sending screenshots, photos of documents, UI mockups, or error screens. Audio and video support depends on how your OpenClaw instance handles file uploads.
If you're just running OpenClaw day-to-day (which is most people), use the slim variant. It's smaller, starts faster, and has a reduced security footprint. Only use the standard image if you're actively developing or contributing to OpenClaw's codebase and need the build tools included.
OPENCLAW_VARIANT=slim in your Docker config to get a smaller, faster, more secure container imagegemini-3.1-flash-lite and OpenClaw knows what you meanTomorrow we're going to dig into model aliases and normalization in OpenClaw — how the system maps friendly model names to actual API endpoints, and how you can create your own custom aliases for models you use frequently. It's one of those quality-of-life features that saves you ten seconds every time you use it, which adds up fast.
Found this helpful? Share it with someone who runs OpenClaw. See you tomorrow.
Discover more content: