Engineer runs OpenCode offline with local Qwen3.5 9B via Ollama
Original title: Coding From the Himalayas With an Offline AI Agent
The walkthrough uses OpenCode 1.19.4, blocks web tools, and cites published Qwen scores rather than measurements from the laptop.
A Thomson Reuters principal engineer describes keeping a coding-agent workflow during Himalayan network and power outages by running OpenCode against Qwen3.5 9B served locally by Ollama. Models, project dependencies, and tests must already be on the laptop; new package downloads and web lookups still need a connection. The project configuration points only at the local Ollama endpoint, asks before edits or shell commands, and blocks web tools. Cited coding and tool-calling scores come from Qwen’s model card, not this machine.
Unconfirmed: Published benchmark scores and the cloud-model agent study do not measure this offline laptop setup.
- Qwen3.5 9B is listed at about 6.6–7.6 GB to download.
- The model card reports LiveCodeBench v6 65.6 and BFCL-V4 66.1 for the 9B model.
- A preprint found OpenCode used about 40 times Goose’s tokens per solved task.
- Those agent comparisons used cloud APIs, not local Qwen3.5 9B.
- Aider is mentioned as another Ollama-compatible option but is not the walkthrough.
- Organizations
- Thomson Reuters
- AI models
- Qwen3.5 9BQwen3.5 4B
From the original article
9 min read at latent.space
How to set up and run an offline coding agent using OpenCode, Ollama and Qwen.
By Rittika Jindal · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.