Urgent.News

What's breaking now, across hundreds of outlets.

AI

OpenCode and a local LLM handle the boring 80% of my coding, so Claude only gets the hard part

There's nothing stopping me from using both of them at once

OpenCode and a local LLM handle the boring 80% of my coding, so Claude only gets the hard part

OpenCode, a free alternative to Claude Code, has been gaining popularity lately as an open-source option for coding tasks. At first, I was skeptical about the idea of replacing Claude with this free alternative, given the impressive capabilities of Claude's models, even the free Sonnet 5.5 model. However, after testing OpenCode and a local model on my RTX 4070 Ti Super, I found that the combination performed surprisingly well.

While I remain cautious about relying solely on local models, I believe that OpenCode and a local model can create a more economical workflow when used together.

For my latest project, a block-breaker game, I used Claude Code (Opus 5.5) on my Pro plan to plan and fix potential issues in the project. The planner wrote 393 lines of code in under two minutes, identifying three challenges that were likely to trip up the local model: collision handling, the paddle's bounce angle, and testing. I then handed the plan to Qwen3.6-35B-A3B inside OpenCode, which read the code, executed the steps, and fixed the issues it identified.

The most impressive aspect of this test was how well Opus predicted the roadblocks. Steps like the F1 key, speed jump between levels, and the high score file were completed successfully on Qwen's first attempt. However, Qwen also made mistakes, such as treating a step that ran without crashing as a successful step, and having its tests occasionally fail due to a lack of context.

Despite these issues, Qwen's overall performance was remarkable, particularly in identifying and fixing bugs that I never would have caught on my own. Overall, it seems that OpenCode and a local model can indeed work together to create a more efficient coding process, with Claude handling the planning and the local model taking care of the execution.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at xda-developers.com →

Report a problem with this story

More in AI

Will AI-Fueled Cyberattacks Lead to Better Funding for Critical Open Source Projects?

Remember that famous XKCD comic joking that all modern digital infrastructure rests on "a project some random person in Nebraska has been thanklessly maintaining since 2003." Politico Europe's…

  • AI-fueled cyberattacks surge, exposing vulnerabilities in PHP, Java, NGINX
  • OpenJS Foundation executive warns team "running on fumes"
  • Palo Alto Networks reports tenfold increase in incidents over four months

Better Recent IPO: SpaceX or Anthropic?

SpaceX and Anthropic pulled off the two largest IPOs ever this year.

  • SpaceX's Q3 revenue surged 161% to $14.6 billion, with $4.4 billion net profit.
  • Anthropic's Q3 revenue grew 1,127% to $17 billion, but recorded $567 million operating loss.
  • Both companies are high-risk, high-reward investments with significant growth potential.

More from Saturday 21 November →