Local AI just got uncomfortably close to the frontier.
Qwen 3.8 27B, Kimi K3, GLM 5.3 and Opus 5 got the exact same game prompt. All four shipped something playable. Opus 5 still won clearly. But Qwen runs locally on a single RTX 4090 at ~14–17 GB in 4-bit, with 262K context.
Five models. Same prompt.
Build a cinematic LUMA landing page, self-contained, no external assets, high-end typography. All five came back looking expensive. Then I opened the files. Grok 4.6, Gemini 3.7 Flash, GLM-5.3 and Opus 5 all ask the browser for Didot. None of them
Four ais were given one orange slice.
Only one created a character with soul. Same slice. Same canvas. Four models racing on a split screen. Kimi K3 drew a basic stick figure with a pencil. GPT 5.6 Sol barely added limbs and a smile. Gemini 3.7 Flash gave it a dress and elegant
Qwen 3.8 27B, Kimi K3, GLM 5.3 and Opus 5 got the exact same game prompt.
You have reached the end of the archive
All of Claude Opus 5