CLI Coding Agent

Basic idea

Use Ollama on DGX Spark to run a local coding model and connect a CLI coding agent. This playbook supports three options: Claude Code, OpenCode, and Codex CLI. Each agent is wired up with Ollama's built-in launch method (ollama launch <agent>), so you can work without environment variables, provider config files, or external cloud APIs.

Choose your CLI agent

Pick the tab that matches the CLI agent you want to use:

• Claude Code: Fastest path to a working CLI agent with a local Ollama model.
• OpenCode: Open-source CLI launched directly from Ollama.
• Codex CLI: OpenAI Codex CLI launched directly from Ollama against the local model.

What you'll accomplish

You will run a local coding model (Qwen3.6) on your DGX Spark with Ollama, launch your chosen CLI agent against it with a single command, and complete a small coding task end-to-end.

What to know before starting

• Comfort with Linux command line basics
• Experience running terminal-based tools and editors
• Familiarity with Python for the short coding task

Prerequisites

• DGX Spark access with NVIDIA DGX OS 7.3.1 (Ubuntu 24.04.3 LTS base)
• Internet access to download model weights
• Ollama v0.15 or newer (required for ollama launch)
• GPU memory depends on the Qwen3.6 variant you choose:
  • qwen3.6:latest (35B-a3b, MoE) — ~24GB, 256K context
  • qwen3.6:35b-a3b-nvfp4 — ~22GB, NVIDIA FP4 build tuned for Blackwell (DGX Spark)
  • qwen3.6:35b-a3b-q8_0 — ~39GB, higher-quality quant
  • qwen3.6:35b-a3b-bf16 — ~71GB, full precision (fits Spark's unified memory)

Time & risk

• Duration: ~15-25 minutes (mostly model download time)
• Risk level: Low
  • Large model downloads can fail if network connectivity is unstable
  • Ollama versions older than 0.15 do not support ollama launch
• Rollback: Stop Ollama and delete the downloaded model from ~/.ollama/models
• Last Updated: 04/16/2026
  • Switched to ollama launch method and upgraded the default model to Qwen3.6