Latent Space published a walkthrough by a Thomson Reuters Principal Engineer who moved to the foothills of the Himalayas and faced regular network and power outages. The piece documents running a coding agent entirely offline on a laptop. The setup combines OpenCode, an open-source coding agent that reads files, proposes edits and runs commands such as test suites with permission prompts; Ollama, which runs the model locally; and Qwen, an open-weight model whose parameters can be downloaded and run on the same machine. The author selected Qwen3.5 9B over the 4B variant, citing stronger published coding and tool-calling results in Qwen's model card on LiveCodeBench v6 and BFCL-V4 benchmarks. The article notes these are scores reported in the model card, not measurements from the author's laptop. On memory, it references Voltti budgeting 12–16 GB for a 6.6 GB Qwen3.5 9B build, and warns a 64k-token OpenCode configuration may need more. The workflow requires downloading the model, dependencies and test tools first, then disconnecting. Tasks like explaining a function, investigating a failing test or refactoring can run offline; installing new packages or web lookups still need a connection. The author also cites a preprint reporting OpenCode used roughly 40 times the tokens of Goose.
- Setup uses OpenCode (agent), Ollama (model runner) and Qwen3.5 9B (open-weight model)
- Voltti budgets 12–16 GB memory for a 6.6 GB Qwen3.5 9B build; 64k-token config may need more
- Offline tasks require model, dependencies and test tools downloaded in advance
- Benchmark scores cited are from Qwen's model card, not measured on the author's laptop
What it means for you
This is a hands-on demonstration that a usable coding assistant can run on a single laptop with no internet, using free, open tools. For most operations teams it changes little day to day, but it's a concrete proof point that AI help doesn't always require sending your code or data to an outside service — which matters if your team works in places with poor connectivity, or handles code that can't leave your machines.
Try this
If anyone on your team already works offline or with sensitive code, have one developer spend an afternoon trying this setup on a spare laptop and report back on speed and usefulness before you draw any conclusions.
Who should care
Engineering leads whose teams handle code or data that can't go to cloud services, and anyone managing developers who work in low-connectivity or air-gapped environments.
Skip this if
Your team always has reliable internet and no restrictions on using cloud-based coding tools — in that case this is a neat experiment, not something to act on.
Sources: Latent Space — read the original