OpenCode + MiniMax M3: setup, demos and model comparisons

The presenter reviews MiniMax M3 in OpenCode, an open source terminal coding agent. The video describes M3 as an open-weight model with a context window of up to one million tokens and native image and video input. It explains MiniMax Sparse Attention and reports the company's claims of lower compute costs and faster processing at long context lengths.

Setup uses hosted access: create an OpenCode Zen account, copy an API key, connect through OpenCode Go and select M3 from the model menu. This walkthrough does not demonstrate how to run models locally. The presenter distinguishes the paid Go plan from a separate free model tier and says M3's free availability is promotional, with no guarantee it will continue.

Two coding demonstrations cover an interactive Rubik's Cube in a single HTML file and a project management dashboard with draggable task cards. The presenter reports working controls and favors M3 for frontend prototyping.

Reported SWE-bench Pro results put M3 at 59%, above GPT-5.5 and Gemini 3.1 Pro but below Claude Opus 4.7. Other cited tests show weaker performance on extended AI agent tasks and computer use. The presenter also notes that the comparisons use Opus 4.7 rather than the newer 4.8. These benchmark figures and demo assessments are claims presented in the video, rather than independent verification.

The video also promotes CodeSnif, a pull-request review service.