Set Up Nvidia's Free Nemotron 3 Ultra for Cost-Saving Coding Agents
Replace paid coding AI tools with a step-by-step guide to configure a free, open-source model with massive context for agentic workflows.
Practical Summary
This tutorial walks you through installing and connecting Nvidia's Nemotron 3 Ultra model—a free, open-source alternative to paid coding AIs—to the OpenCode tool in five steps, enabling zero-cost, long-context coding agent workflows.
Why It Matters
It demonstrates a repeatable workflow to eliminate subscription costs for coding AI tools, with evidence of free access and practical steps, while honestly noting the trade-offs in rate limits and performance.
Step 1: Get a Free API Key
Visit the OpenRouter website (https://openrouter.ai/), sign in with Google or GitHub, navigate to the keys page, and create a new API key. Copy it for later use. The model 'nvidia/nemotron-3-ultra-550b-a55b:free' is listed at $0 cost with no credit card required, though it is rate-limited. As an alternative, Nvidia's developer program also offers free prototyping access.
Step 2: Install OpenCode
Run the one-liner command in your terminal to install OpenCode: 'curl -fsSL https://opencode.ai/install | bash'. After installation, verify it by running 'opencode --version'. This tool does not require Docker.
Step 3: Connect OpenRouter to OpenCode
Launch OpenCode by running 'opencode' in your terminal. In its text-based interface (TUI), type '/connect', search for 'openrouter', and paste the API key you copied earlier. The connection details will be stored automatically in a local file (~/.local/share/opencode/auth.json).
Step 4: Select the Nemotron Model
In the OpenCode TUI, use the '/models' command to list available models. Look for and select 'nvidia/nemotron-3-ultra-550b-a55b:free' from the list. If it's not pre-loaded, manually add it by editing the configuration file at '~/.config/opencode/opencode.json' with the provided JSON snippet, then restart OpenCode.
Step 5: Start Your Coding Workflow
Change directory into your code repository, launch OpenCode, and begin prompting the model. The 1M token context window is suited for long, tool-heavy agentic sessions. Note that the free tier is rate-limited and best for testing, not continuous production, and the model may lag behind top closed models in raw intelligence benchmarks.