How to use Unknown to test a local coding model
The author tested local Qwen through OpenCode on three coding tasks. An MCP feature completed without human input, a Nix package port looped on a hash, and a harder task was stopped after repeated discovery and context compactions.
news.ycombinator.comPrompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Use [OpenCode configuration] with [local model] to implement [bounded MCP feature]. Compare the result and human interventions against [reference model]. Then attempt [Nix package port] and [harder task], recording completion, discovery loops and context compactions. Report incomplete work explicitly.
Reach people exploring AI agents
See placements →What happened
The MCP feature completed in 40 minutes without author input, with results described as nearly identical to a 20-minute Sonnet-5 attempt. Later tasks looped or were stopped.
Reported by 0x457 on Hacker NewsHow the agent did it
- Compared streamed responses from multiple models through a knowledge database workflow.
- Configured OpenCode to use local Qwen and assigned an MCP feature task.
- Tried porting a package version into a Nix flake and observed a hash-discovery loop.
Where a person stepped in
The author compared results with another model, observed failure loops and stopped the harder run when the room became too hot.
A person approved each step before the agent continued.
The author acquired an R9700 to run Qwen 3.8-27B locally and compared model responses through a knowledge database workflow. That setup queried multiple models, streamed their answers and let the author choose a result. They report preferring Qwen over the other contenders, which had been selected before the hardware upgrade and were smaller models.
They then connected OpenCode to the local model and tried three coding tasks. A new feature for their MCP completed in forty minutes with no input from the author. They described the result as nearly identical to a Sonnet-5 attempt that took twenty minutes and required repeated interaction.
The next task, porting a specific package version into a Nix flake, stalled in a loop around finding a hash that Nix had already supplied. A harder third task repeatedly performed discovery and exhausted context, passing through three compactions. The author stopped it when the room became too hot. The account therefore includes both a completed task and unresolved attempts, rather than a uniformly successful local coding workflow. Who it's for: developers evaluating local coding models who can compare outputs and monitor failures during experimental tasks.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · 0x457
Frequently asked questions
Did every local coding task complete?
No. The MCP feature completed, but the Nix package port looped around obtaining a hash. A harder task repeatedly performed discovery and was stopped after three context compactions.
How did the first task compare with Sonnet-5?
The author reports 40 minutes with zero input for local Qwen versus 20 minutes for Sonnet-5, which kept requiring interaction. They described the results as nearly identical.
More details
Remaining steps
- Assigned a harder task and stopped it after repeated discovery, context compactions and room heating.
About this account
The first OpenCode task completed with zero author input; later tasks looped or exhausted context and one was stopped by the author.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=49391429. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: 0x457 on Hacker News
Source published 21 Aug 2026
Source checked 14 Oct 2026
Tools and topics
OpenCode · Qwen 3.8-27B · R9700 · Nix
local models · coding evaluation · Nix
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.